Ollama: baixar versions anteriors

Speech
Ollama can now generate speech on macOS. ollama run plays text or a chat model's reply aloud as the audio is generated, and apps can stream audio through the OpenAI-compatible /v1/audio/speech endpoint.
ollama run phonon "Hello from Ollama."
d3 decision models
d3 brings vLLM Semantic Router's decision models to Ollama. Give a model text, structured state, or an image, and ask it to classify, answer yes or no, or score options with probabilities. The family includes 2B, 4B, 9B, and 27B sizes.
ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"
New Contributors
- @SynapShift made their first contribution in https://github.com/ollama/ollama/pull/19173
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.10...v0.40.11-rc0

NVIDIA Nemotron 3.5 Bolt E4B
Nemotron 3.5 Bolt E4B is NVIDIA’s smaller, Gemma 4-based vision-language model for fast local conversations, image understanding, and tool use. It supports a 128K-token context window.
The model includes a DFlash draft for speculative decoding.
ollama run nemotron-3.5-bolt:e4b
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.9...v0.40.10

Speech
Ollama can now generate speech with Phonon on Apple Silicon. ollama run plays text or a chat model's reply aloud as the audio is generated, and apps can stream audio through the OpenAI-compatible /v1/audio/speech endpoint.
ollama run phonon "Hello from Ollama."
d3 decision models
d3 brings vLLM Semantic Router's decision models to Ollama. Give a model text, structured state, or an image, and ask it to classify, answer yes or no, or score options with probabilities. The family includes 2B, 4B, 9B, and 27B sizes.
ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"
Flint on Linux and Windows
Flint now runs on Linux and Windows as well as macOS.
Other improvements
- Streaming OpenAI Responses and Anthropic Messages requests now report generation failures instead of ending silently after partial output.
- ollama launch now passes cloud models' context lengths and capabilities to coding tools, fixing incorrect context and compaction settings.
- Anthropic-compatible web search now uses the model's supported thinking setting in follow-up requests, preventing avoidable failures.
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.8...v0.40.9

Recursive Flint
Flint brings Recursive's coding model to Ollama. It is designed for the work coding agents do across a repository: understanding unfamiliar code, tracing bugs, implementing features, and adding tests. Reasoning, multi-turn tool use, and a 262K-token context window support longer tasks across files and tool interactions.
Flint is an Apache 2.0-licensed, 40B-parameter mixture-of-experts model with approximately 2.5B parameters active per token. In its release evaluations, Recursive reports 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified.
ollama run flint
Note: Flint is supported on MacOS. Other GPU support will be enabled in a subsequent release.
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.7...v0.40.8

What's Changed
- pplx-decider is a 27B decision model from Perplexity, fine-tuned from
Qwen3.8-27B
- Continuing an OpenAI-compatible Responses conversation no longer fails when switching between models with and without
thinking, or when the history contains another provider's compaction items
- Fixed streaming Responses requests where client tool calls (e.g. Codex with web search) could emit before open
reasoning items, producing misordered/duplicate tool call events
- Requests routed to remote models are now cancelled on ollama.com when the client disconnects, instead of continuing
to run
- ollama show now reports which renderer and parser a model uses and where its chat template comes from; the API
exposes the same values in new renderer, parser, and template_source fields
- More draft models now work with speculative decoding: llama-server infers the draft type automatically instead of
requiring Ollama to identify it
- ollama launch claude now clamps CLAUDE_CODE_AUTO_COMPACT_WINDOW to Claude Code's supported range, avoiding a
"window too large" warning
- Updated llama.cpp and MLX
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.6...v0.40.7

What's Changed
- win: prevent orphaned runners by @dhiltgen in https://github.com/ollama/ollama/pull/18500
- gguf: add the q2_0 tensor and file types by @jmorganca in https://github.com/ollama/ollama/pull/18967
- Launch clef in Ollama's cloud by @jberg5 in https://github.com/ollama/ollama/pull/19036
- use identifiers for --ask mode question keys by @jberg5 in https://github.com/ollama/ollama/pull/19041
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.5...v0.40.6-rc0

NVIDIA Nemotron 3.5 Super
NVIDIA Nemotron 3.5 Super is a vision-language model for coding, reasoning, tool use, and document understanding. It pairs a 120B-parameter hybrid Mamba-2/Transformer mixture-of-experts backbone (12B active per token) with a C-RADIO vision encoder, and supports tool calling, an on/off thinking mode, and context up to 1M tokens.
ollama run nemotron-3.5-super
A compact sibling, Nemotron 3.5 Bolt, distills Super down to 74B parameters (9B active) tuned to fit a 24 GB NVIDIA GPU for everyday agentic tasks, with speculative decoding via its included drafter.
What's Changed
- Decision models are now available on cloud — use clef:cloud and clef-flash:cloud to run decision models through ollama.com cloud without pulling them locally
- Continuing an OpenAI-compatible Responses conversation that includes custom tool calls no longer fails with an error
- Clearer error messages when cloud access fails: sign-in problems, payment issues in the desktop app, and upstream access errors from the ChatGPT gateway now explain what went wrong
- Updated llama.cpp (adding Nemotron 3.5 support) and MLX
New Contributors
- @willietran made their first contribution in https://github.com/ollama/ollama/pull/18986
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.4...v0.40.5

Decision models in the CLI
Decision models can now be used directly from the terminal with ollama run — ask yes/no questions, classify text into labels, or score against a rubric:
ollama run nimble --ask "Is a refund being requested?" < ticket.txt
What's Changed
- Kev decision models are now supported by the MLX runner on Apple silicon
- ollama launch can now be used with usage credits without upgrading to a paid plan
- ollama launch codex now includes locally installed Ollama models in its model picker. Use /model within the Codex CLI to swap models.
- Improved inference performance by reusing verified manifests and already loaded models
- Added support for multimodal embeddings through llama-server
- Fixed loading multimodal embedding models with small context windows
- Improved Gemma 4 prefill performance on NVIDIA GPUs with the MLX runner
- MLX runner now clips long Laya states instead of rejecting them
- Improved speculative decoding reliability in the MLX runner, including recurrent state handling after rollback
- Updated llama.cpp with native decision model support
- Updated README and documentation with clearer privacy and cloud-services wording
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.3...v0.40.4

What's Changed
- ollama launch claude no longer shows a warning about Claude connectors by @hoyyeva in https://github.com/ollama/ollama/pull/18878
- ollama launch codex no longer warns about service tiers by @hoyyeva in https://github.com/ollama/ollama/pull/18879
- Paused background model upgrade for embedding and other request performance, by @jmorganca in https://github.com/ollama/ollama/pull/18908
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.2...v0.40.3

Model upgrades
Models downloaded with earlier versions of Ollama are upgraded in the background the first time you run them, for better performance and compatibility when running on llama.cpp.
To make downgrading safe, Ollama keeps the original copy as a backup, so upgraded models are temporarily kept on disk. A future release will remove these backups automatically.
To remove backed up models now (requires jq):
curl -s localhost:11434/api/tags | jq -r '.models[].name' | while read -r m; do
curl -s localhost:11434/api/show -d "{\"model\":\"$m\"}" | jq -r \
'select(any(.manifests[]?; .runner == "llamacpp")) | .manifests[] | select(.runner == "ggml") | .digest'
done | sort -u | xargs -n1 ollama rm
This only deletes the backups. If you later downgrade to a version older than 0.40, you'll need to re-pull those models.
Other changes
- ollama list no longer shows duplicate entries for upgraded models by @dhiltgen in https://github.com/ollama/ollama/pull/18874
- ollama launch claude uses the model's full context length by @jberg5 in https://github.com/ollama/ollama/pull/18855
- README: add oxi to community integrations by @maziluiosif in https://github.com/ollama/ollama/pull/18739
New Contributors
- @maziluiosif made their first contribution in https://github.com/ollama/ollama/pull/18739
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.1...v0.40.2-rc0

What's Changed
- server: proxy cloud usage and balance APIs by @drifkin in https://github.com/ollama/ollama/pull/18829
- llama: fix clef head reads past 2GiB on windows by @Gigrise in https://github.com/ollama/ollama/pull/18777
- cmd: remove account step from CLI onboarding by @hoyyeva in https://github.com/ollama/ollama/pull/18826
- manifest: avoid symlinks on Windows by @dhiltgen in https://github.com/ollama/ollama/pull/18852
- docs: fix 6 dead links in README community integrations list by @aniketkrs in https://github.com/ollama/ollama/pull/18814
- docs: fix broken download links in app README by @chenlichao in https://github.com/ollama/ollama/pull/18233
- mlx: drop carried metal residency patch now that it is upstream by @dhiltgen in https://github.com/ollama/ollama/pull/18854
New Contributors
- @Gigrise made their first contribution in https://github.com/ollama/ollama/pull/18777
- @aniketkrs made their first contribution in https://github.com/ollama/ollama/pull/18814
- @chenlichao made their first contribution in https://github.com/ollama/ollama/pull/18233
Full Changelog: https://github.com/ollama/ollama/compare/v0.40.0...v0.40.1-rc0

What's Changed
Models run on MLX on Apple Silicon by default
In this release, on Apple Silicon devices, model architectures supported by the MLX runtime will automatically run on MLX.
ollama pull qwen3.8
ollama run qwen3.8
Additional models include gemma4, qwen3.6 and qwen3.5
Decision models are now available on MLX as well: Nimble tev1 clef clef-flash
MLX now has support for an embedding model: embeddinggemma-2
We will continue testing and enabling additional models.
Full Changelog: https://github.com/ollama/ollama/compare/v0.35.1...v0.40.0

Clef decision models
Ollama now supports Clef and Clef Flash, Cloudflare's new open-source decision models, through /v1/systemone.
Clef (27B) and Clef Flash (9B) are multimodal: requests can now include images alongside the text state, shared by all questions and scored jointly with it.
curl http://localhost:11434/v1/systemone -d '{
"model": "clef-flash",
"state": "The user took this screenshot.",
"images": ["<base64-encoded image>"],
"questions": {
"has_ollama": {"type": "noul", "instructions": "Does this image contain Ollama?"}
}
}'
{
"model": "clef-flash",
"answers": {
"has_ollama": {
"type": "noul",
"noul": 0.958
}
},
"usage": {
"input_tokens": 548,
"output_tokens": 0
}
}
What's Changed
- Models using web search can now perform up to ten searches per response, up from three
- Modelfiles now support CAPABILITY declarations, so model creators can explicitly declare what a model can do. Declarations are preserved when creating from GGUF or safetensors, through model inheritance, and on Modelfile export
- ollama show and the model list now report only decision as the capability for decision models, so clients no longer offer them for general chat, tools, or thinking
- Updated llama.cpp and the MLX engine
Full Changelog: https://github.com/ollama/ollama/compare/v0.35.0...v0.35.1

Decision models
Ollama now supports decision models through /v1/systemone, based on TypeSafe’s Jev API.
Decision models return choices, probabilities, and scores instead of text. Use them for tasks such as ticket triage, model routing, and content classification.
Available models:
- Nimble from Bespoke Labs
- Tev1 from Together AI
ollama pull nimble
Send context and one or more questions:
curl http://localhost:11434/v1/systemone \
-H 'Content-Type: application/json' \
-d '{
"model": "nimble",
"state": "Our checkout has returned 500 errors since 9am.",
"questions": {
"label": {
"type": "choice",
"instructions": "Which label fits this ticket?",
"criteria": {
"billing": "Payments and refunds",
"bug": "Software errors",
"account": "Login and account access"
}
}
}
}'
Example response:
{
"model": "nimble",
"answers": {
"label": {
"type": "choice",
"choice": "bug",
"probabilities": {
"billing": 0.0125,
"bug": 0.9781,
"account": 0.0093
},
"confidence": 0.8906
}
},
"usage": {
"input_tokens": 174,
"output_tokens": 1
}
}
The API supports three question types:
- choice: Select an option and return probabilities for each.
- noul: Return the probability that a condition is true.
- score: Return a score across an ordered set of criteria
What's Changed
- Settings now opens without waiting for model discovery.
- Fixed the macOS update menu and icon not reflecting an available update at startup.
- Fixed stalled MLX model downloads hanging indefinitely.
- Requests containing the deprecated typical_p parameter now log a warning instead of failing.
Full Changelog: https://github.com/ollama/ollama/compare/v0.34.4...v0.35.0

What's Changed
- Structured outputs on thinking models now apply in a single pass, making them faster and more reliable.
- Fixed intermittent "model not found" errors with a large local library
- Fixed the macOS app becoming unresponsive when checking if ChatGPT or Codex is running.
- Qwen 3.8 prompt processing is faster on Apple Silicon.
- Gemma 4 on Apple Silicon now picks the best image resolution per image, keeping more detail in high-resolution images.
- Updated llama.cpp, MLX, and XGrammar.
Full Changelog: https://github.com/ollama/ollama/compare/v0.34.3...v0.34.4

What's Changed
GET /api/show now advertises each model's thinking controls and default:
Available in the CLI with:
ollama show gemma4
thinking
levels false, true
default true
Available in the API with:
curl http://localhost:11434/api/show -d '{"model": "glm-5.3-flash:cloud"}'
{
"thinking": {
"values": ["low", "high", "max"],
"default": "max"
}
}
Also available on ollama.com directly for cloud models.
- Nemotron H vision models are now supported on Apple Silicon with MLX
- Ollama's macOS app will now no longer reopen windows you've closed when activating it
- Fix for model pulls from HuggingFace
Full Changelog: https://github.com/ollama/ollama/compare/v0.34.2...v0.34.3

What's Changed
- Added first-run setup when running ollama, with options to sign in or continue locally. Setup completion is shared with the desktop app on macOS and Windows.
- Added ollama://apps to open the desktop app’s Apps page directly on macOS and Windows.
- Fixed excessive memory growth during long generations with MLX speculative decoding.
- Updated llama.cpp.
Full Changelog: https://github.com/ollama/ollama/compare/v0.34.1...v0.34.2

What's Changed
- MLX safetensors ollama create no longer experimental. GGUF model creation now requires using llama.cpp tooling for safetensor conversion and quantization.
- Improved MLX memory handling on Apple Silicon
- Runaway repeat token detection now requires 100 repeat tokens for reduced false positives (e.g. OCR)
- /api/tags is much faster on large model libraries (3.1 s → 294 ms cold in testing), and model capabilities are now reported consistently.
- Deprecated typical_p: it can no longer be set when creating new models, existing GGUF models retain support.
- MLX and llama.cpp updates
Full Changelog: https://github.com/ollama/ollama/compare/v0.34.0...v0.34.1-rc1

Use Ollama models in ChatGPT Desktop
Ollama models can now be used directly in ChatGPT Desktop, so you can keep your existing workflow while running open models. Setup is available from the Ollama app on MacOS.
This release also improves structured output performance on Apple Silicon, adds support for OpenAI-compatible client tool search and response compaction.
Full Changelog: https://github.com/ollama/ollama/compare/v0.33.3...v0.34.0

What's Changed
- gemma4 now supports images and audio on MLX engine
- Report cached prompt tokens
- Honor GGUF model defined default parameters
- MLX, MLX-C, llama.cpp update
New Contributors
- @marcelpetrick made their first contribution in https://github.com/ollama/ollama/pull/17579
Full Changelog: https://github.com/ollama/ollama/compare/v0.33.2...v0.33.3

What's Changed
- Ollama's app now follows the system appearance again, restoring dark mode support
- Fixed the macOS app to properly hand off to an already-running instance instead of starting a second one
- The Claude Desktop proxy no longer interrupts in-flight requests when the model catalog updates
Full Changelog: https://github.com/ollama/ollama/compare/v0.33.1...v0.33.2

What's Changed
- MLX: Qwen3.8 Flash Next support
- cmake: make external compat patches idempotent
- MLX and llama.cpp update
- mlxrunner: add structured output support
- mlxrunner: avoid Metal GPU timeouts when loading models from slow storage
New Contributors
- @pd95 made their first contribution in https://github.com/ollama/ollama/pull/17948
Full Changelog: https://github.com/ollama/ollama/compare/v0.33.0...v0.33.1

What's Changed
Claude Desktop
Developers can now easily configure Claude Desktop to seamlessly work with Ollama as a third-party gateway provider.
Improved caching
- Fixed a hang where agent clients that cancel long prefills
- Prefill restore points are now trustworthy by construction: a cancelled prefill keeps every restore point it crossed, so retries resume where they stopped instead of restarting from scratch
- Resumed prefills no longer record restore points that fail to cover what they claim; on models with recurrent layers this previously forced a request matching 46k of 47k tokens to reprocess from zero
- Disabled Claude Code's "tokens left" token-countdown system message, which Ollama moved to the front of the prompt and broke the KV cache on every request
Other improvements
- DeepSeek Harness launcher now falls back to npx when the global npm install fails, with Windows command-shim support
- MLX dependency update (#17886)
- Fixed broken default packaging caused by macOS-specific assumptions affecting Linux/Windows builds
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.15...v0.33.0

What's Changed
- New desktop onboarding flow on first launch
- Caches resolved model metadata between requests, cutting time-to-first-token by roughly half (TTFT dropped from ~995 ms to ~524 ms in benchmarks)
- Fixes a bug where chat and generate could wedge after a mid-stream parser error
- Qwen 3.8 system messages are now normalized so non-leading system messages are handled consistently
- MLX and llama.cpp dependency updates
New Contributors
- @gaugarg-nv made their first contribution in https://github.com/ollama/ollama/pull/17752
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.14...v0.32.15

What's Changed
- llm: transcode WebP images for llama-server
- renderers/qwen: tolerate non-leading system messages
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.13...v0.32.14

What's Changed
- qwen3.8: support developer instructions
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.12...v0.32.13

Qwen 3.8 - 27B model support
This release adds the support of Qwen 3.8 27B.
For Apple Silicon devices, Ollama has in particular optimized for maximum performance and output quality suitable for repeated tasks and coding agents.
Qwen 3.8 model page:
https://ollama.com/library/qwen3.8
Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks.

What's Changed
- ollama launch dsh now supports DeepSeek Harness, DeepSeek's open-source agent harness
- ollama launch muse now supports Muse Code, Meta's agentic coding CLI
- The OpenAI-compatible Responses API now supports web search
- Muse Glimmer template updates
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.10...v0.32.11

NVIDIA Nemotron 3.5 Lightning
NVIDIA Nemotron 3.5 Lightning is an open 30B mixture-of-experts (MoE) model with 3B active parameters built for that execution layer of always-on agents. It is designed for harnesses like OpenClaw and Hermes Agent – all supported by the NVIDIA NemoClaw open source security and management stack for running always-on AI agents.
ollama run nemotron-3.5-lightning
What's Changed
- Added the Nemotron 3 architecture
- Handle boundary condition in Muse Glimmer function calling parser
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.8...v0.32.9

What's Changed
- Add Muse Glimmer support for NVIDIA, AMD, and additional platforms
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.7...v0.32.8-rc0

Muse Glimmer
Note: Muse Glimmer is currently available via initial support via Ollama's MLX engine on Apple Silicon. Support for NVIDIA, AMD, and other platforms will be available in the coming days.
Muse Glimmer, Meta's newest open model and the first released by Meta Superintelligence Labs, is now available on Ollama. It's a 30B multimodal model purpose-built for agent workloads that run locally.
With Ollama, you can now use Muse Glimmer to power coding agent applications such as Claude Code, Codex, Pi and more, as well as long-running personal assistants such as OpenClaw and Hermes.
Ollama's MLX engine provides state-of-the-art performance on Apple Silicon for this model, with support for DFlash and image input as of Ollama 0.32.7.
To download and run Muse Glimmer locally:
ollama run muse-glimmer:30b-mlx
To run Muse Glimmer on Apple Silicon with Claude Code, download Ollama and run:
ollama launch claude --model muse-glimmer:30b-mlx
For a lighter-weight coding agent, try Pi:
ollama launch pi --model muse-glimmer:30b-mlx
For personal assistant frameworks such as OpenClaw and Hermes, use:
ollama launch openclaw --model muse-glimmer:30b-mlx
ollama launch hermes --model muse-glimmer:30b-mlx

What's Changed
- Qwen3.5 is faster on Apple GPUs: the MLX engine now uses the model's MTP head for speculative decoding automatically
- /v1/chat/completions streaming now matches OpenAI's wire format: role only on the first chunk, finish_reason on its own chunk,
and usage in a separate chunk with stream_options.include_usage.
- Truncated OpenAI responses now report finish_reason: "length" instead of "tool_calls".
- ollama run kimi-k3 now offers kimi-k3:cloud for cloud-only models that publish no default tag, instead of failing.
- TUI fixes: pipe-delimited prose no longer renders as a table, Enter accepts the highlighted @ file completion, and /prompt
scrolling is no longer laggy.
- Experimental image generation has been temporarily removed. Continue using 0.32.5 for image generation support
- Updated the MLX and llama.cpp engines.
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.5...v0.32.6-rc0

What's Changed
- Fixed an MLX Metal bug that could reduce output quality for NVFP4 models, particularly Laguna.
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.4...v0.32.5

What's Changed
- Support Laguna on Apple GPUs via the MLX engine
- Quantize draft-model output heads at the requested type when creating speculative-decoding drafts.
- Fixed Qwen3 MoE decoding for differently-quantized experts, plus faster packed gate/up projection (~4–9% on M5 Max).
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.3...v0.32.4

What's Changed
- Fixed model downloads that stall before sending data.
- Improved integrations: restored Claude Code Channels, fixed Anthropic thinking streams, and made Hermes Desktop respect --force-build.
- Expanded GPU support with CUDA on Windows ARM64, B200 support through CUDA 12, and lower memory use on Linux CUDA/ROCm iGPUs.
- Added chat, thinking, and tool calling support for Laguna 2.1 models, including a Metal inference fix.
- Fixed GLM tool calls being silently dropped at the end of generation.
- Updated the MLX and llama.cpp engines.
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.1...v0.32.3

What's Changed
- Improved Gemma 4 tool calling and multi-turn reasoning, including more reliable tool-response continuations
- Fixed a recurrent MLX model cache leak that could increase memory use across requests, and improved cache snapshot performance
- MLX text model loading now respects OLLAMA_LOAD_TIMEOUT
- Agent web search and fetch now tell users to run ollama signin when authentication is required
- The interactive agent now receives the current working directory for better project context
- Fixed ollama launch so choosing Pick another model for a deprecated model passed with --model opens the model picker
- Updated VS Code setup documentation for the official Ollama extension
Full Changelog: https://github.com/ollama/ollama/compare/v0.32.0...v0.32.1-rc0

What's Changed
- New interactive agent experience: running ollama now launches an agent to help you code and delegate work
❯ ollama
Ollama 0.32.0

▸ Chat, Code, & Work (glm-5.2:cloud)
Chat with models, code, search the web, and delegate real work
- Renamed the Codex App integration to ChatGPT: use ollama launch chatgpt (and --restore to return to your usual ChatGPT profile)
- Simplified integration selection: the ollama launch menu now only offers the most popular integrations (other integrations can be accessed through ollama launch
- Warns before launching older agent models: CodeLlama, Qwen2.5(-coder), Llama 3.x, Mistral, StarCoder, and the base DeepSeek-R1 tags now prompt a deprecation warning before ollama launch continues
Full Changelog: https://github.com/ollama/ollama/compare/v0.31.2...v0.32.0

What's Changed
- Enabled flash attention on older NVIDIA GPUs (compute capability 6.x)
- iGPU can now offload vision models with padding to fit available memory
- Fixed structured output for thinking models when thinking is disabled
- Hardened GGUF model creation
- ollama launch for Claude Code now disables telemetry by default
- Fixed loading models on paths with non-UTF-8 characters
- Updated the MLX and llama.cpp engines
New Contributors
- @kevinpark1217 made their first contribution in https://github.com/ollama/ollama/pull/16949
Full Changelog: https://github.com/ollama/ollama/compare/v0.31.1...v0.31.2

What's Changed
- mlx: tighten up gemma4 moe loading code by @pdevine in https://github.com/ollama/ollama/pull/16964
- mlx: bump to latest version to include new small batch matmul kernel @jessegross @dhiltgen
- llama.cpp: bump to b9840 @dhiltgen
- improved gemma4 MTP performance @jessegross
Full Changelog: https://github.com/ollama/ollama/compare/v0.31.0...v0.31.1

What's Changed
- launch: add thinking capability detection to opencode by @hoyyeva in https://github.com/ollama/ollama/pull/15434
- launch: auto-install Claude Code by @hoyyeva in https://github.com/ollama/ollama/pull/16802
- launch: auto-install opencode when missing by @hoyyeva in https://github.com/ollama/ollama/pull/16806
- discover: fix inverted iGPU/dGPU Vulkan classification on Windows hybrid graphics by @Sahil170595 in https://github.com/ollama/ollama/pull/16669
- mlxrunner: unify and tune speculative decoding by @jessegross in https://github.com/ollama/ollama/pull/16791
- launch/codex: detect model drift when Codex App UI switches by @BruceMacD in https://github.com/ollama/ollama/pull/16864
- llama: add sm_86 architecture to cuda_v13_windows preset by @anishesg in https://github.com/ollama/ollama/pull/16834
- llm: size mmproj offload by projector memory by @dhiltgen in https://github.com/ollama/ollama/pull/16866
- docs: document max think level by @ParthSareen in https://github.com/ollama/ollama/pull/16877
- llm: preserve generation headroom for shifted prompts by @ParthSareen in https://github.com/ollama/ollama/pull/16856
- llama: default qwen2.5vl window attention metadata by @dhiltgen in https://github.com/ollama/ollama/pull/16868
- llm: use host Vulkan loader on Windows by @dhiltgen in https://github.com/ollama/ollama/pull/16869
- mlx: update and fix CUDA JIT packaging by @dhiltgen in https://github.com/ollama/ollama/pull/16871
- llm: fix ollama ps double-counting mmap'd weights on partial offload by @discobot in https://github.com/ollama/ollama/pull/16709
- docs: redesign docs landing and integrations overview by @hoyyeva in https://github.com/ollama/ollama/pull/16807
- server: align generate with native chat templates by @dhiltgen in https://github.com/ollama/ollama/pull/16878
- jetson: add CC 87 for CUDA v13 by @dhiltgen in https://github.com/ollama/ollama/pull/16628
- llama.cpp version update by @dhiltgen in https://github.com/ollama/ollama/pull/16548
New Contributors
- @Sahil170595 made their first contribution in https://github.com/ollama/ollama/pull/16669
- @anishesg made their first contribution in https://github.com/ollama/ollama/pull/16834
- @discobot made their first contribution in https://github.com/ollama/ollama/pull/16709
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.10...v0.30.11-rc0

What's Changed
- models: add Cohere2MoE model by @jmorganca in https://github.com/ollama/ollama/pull/16670
- llama: update llama.cpp to b9672 by @pdevine in https://github.com/ollama/ollama/pull/16775
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.9...v0.30.10-rc0

What's Changed
- Support for Cohere2Moe architecture
- Fixed LFM2 parser/render for cases where thinking was not emitted
- Fixed issue where ollama launch claude and other coding agent or assistant use cases would only output one token
- Ollama will now return an error if a single message is larger than the current context window
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.8...v0.30.9-rc1

What's Changed
- Fixed ollama launch selecting the wrong provider in some cases
- Improved prompt caching by decoupling it from context shift for better KV cache reuse
- More stable MLX inference with hardened linear and embedding layers
- MLX runner now creates snapshots during prompt processing and speculative decoding for improved reliability
- Improved recurrent model support with per-boundary states from the gated-delta kernels
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.7...v0.30.8

Ollama Launch now supports Hermes Desktop, a native desktop interface for the Hermes agent. Run it alongside your Hermes agent to get a visual interface for managing conversations, integrations, and messaging apps.
ollama launch hermes-desktop
What's Changed
- Hermes Desktop is now available via ollama launch hermes-desktop with native Windows configuration path support
- OpenAI-compatible API models list now aligns with available model tags
- Added documentation describing the llama.cpp update process
- Updated Zod schema examples to use the native toJSONSchema helper
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.6...v0.30.7

What's Changed
- ollama launch omp now integrates with Oh My Pi, an AI coding agent with IDE integration
- MLX embedding layers now use NVFP4 global scale for improved quantization on Apple Silicon
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.5...v0.30.6

What's Changed
- llama.cpp version update by @dhiltgen in https://github.com/ollama/ollama/pull/16463
- Kill llama-server during Windows cleanup by @dhiltgen in https://github.com/ollama/ollama/pull/16458
Known Issues
- gemma4:12b crash with floating point exception
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.3...v0.30.4

What's Changed
- models: add support for gemma4-12b by @pdevine in https://github.com/ollama/ollama/pull/16457
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.2...v0.30.3

What's Changed
- feat(launch): show and auto-install Cline CLI by @hoyyeva in https://github.com/ollama/ollama/pull/16402
- log template details to aid troubleshooting by @dhiltgen in https://github.com/ollama/ollama/pull/16403
- cmd/launch: add Qwen code integration by @hoyyeva in https://github.com/ollama/ollama/pull/15900
- launch: fix opencode local model limits by @dhiltgen in https://github.com/ollama/ollama/pull/16425
- llm: include cached prompt tokens in llama-server counts by @dhiltgen in https://github.com/ollama/ollama/pull/16428
- Harden app markdown URL handling by @dhiltgen in https://github.com/ollama/ollama/pull/16380
- discover: allow Radeon 8060S iGPU by default by @dhiltgen in https://github.com/ollama/ollama/pull/16429
- llm: detect llama-server load stalls from output by @dhiltgen in https://github.com/ollama/ollama/pull/16427
- More harden app markdown URL handling by @dhiltgen in https://github.com/ollama/ollama/pull/16436
- llama.cpp version update by @dhiltgen in https://github.com/ollama/ollama/pull/16426
- launch: isolate Codex launch configuration by @ParthSareen in https://github.com/ollama/ollama/pull/16437
- llama: add laguna (poolside) arch via a llama.cpp patch under llama/c… by @dhiltgen in https://github.com/ollama/ollama/pull/16396
- docs: configure hermes desktop app by @BruceMacD in https://github.com/ollama/ollama/pull/16440
- llm: ignore llama-server SSE ping comments by @dhiltgen in https://github.com/ollama/ollama/pull/16443
- fix laguna patch build breakage by @dhiltgen in https://github.com/ollama/ollama/pull/16445
Full Changelog: https://github.com/ollama/ollama/compare/v0.30.0...v0.30.2-rc0

Ollama 0.30 is now available, with improved compatibility and performance using llama.cpp. This augments the MLX engine on Apple Silicon, bringing support to a wider range of hardware.
This release brings support for a wider range of models, including GGUF-based models from Hugging Face and your own fine-tuned models along with faster performance on NVIDIA hardware.
Known issues:
- laguna-xs.2 is not yet supported on Windows/Linux.
- llama3.2-vision is not yet supported
- nomic-embed-text now converts inputs to lowercase per the model card where prior Ollama versions incorrectly preserved mixed case

Codex App
The OpenAI Codex App is now available on Ollama. Use any Ollama model — local or cloud — inside the desktop app to code, browse, and review.
ollama launch codex-app
Built-in browser
Codex can spin up local servers and sites in its built-in browser. Annotate directly on the page to request changes.
Review mode
Review code inside the app, leave comments, and iterate without leaving your workspace.
Choosing a model
For difficult coding and agentic tasks:
- kimi-k2.6 (with vision support)
- glm-5.1
For local use without an Ollama Cloud subscription:
- nemotron-3-super
- gemma4:31b
- qwen3.6
Restore anytime
To switch back to OpenAI's default setup:
ollama launch codex-app --restore
What's Changed
- Reworked the MLX sampler for improved generation quality on Apple Silicon
Full Changelog: https://github.com/ollama/ollama/compare/v0.23.0...v0.24.0

What's Changed
- ollama launch opencode now supports vision models with image inputs
- Fixed formatting of Claude tool results when using local image paths
Full Changelog: https://github.com/ollama/ollama/compare/v0.23.3...v0.23.4

What's Changed
- mlx: refined model push behavior by @dhiltgen in https://github.com/ollama/ollama/pull/15431
- test: integration test hardening by @dhiltgen in https://github.com/ollama/ollama/pull/13532
- app: harden update flows by @dhiltgen in https://github.com/ollama/ollama/pull/16100
- mlx: update the imagegen runner for mlx thread affinity by @pdevine in https://github.com/ollama/ollama/pull/16096
- mlx: avoid status timeout during inference by @dhiltgen in https://github.com/ollama/ollama/pull/16086
- mlx: fix macOS 26 target leakage in v3 metallib by @dhiltgen in https://github.com/ollama/ollama/pull/16053
Full Changelog: https://github.com/ollama/ollama/compare/v0.23.2...v0.23.3

What's Changed
- ollama launch no longer includes Claude Desktop due to the third-party integration being limited to Anthropic models.
- Use ollama launch claude-desktop --restore to restore Claude Desktop to its normal state.
- /api/show responses are now cached, improving median latency by ~6.7x which will increase load speed for integrations like VS Code.
- Improved backup workflow when managing launch integrations
- Cleaner image generation layout in the MLX runner
Full Changelog: https://github.com/ollama/ollama/compare/v0.23.1...v0.23.2

Gemma 4 MTP (Multi-token Processing) for the MLX runner
Gemma 4 MTP speculative decoding is now supported on Macs. This can give over a 2x speed increase for the Gemma 4 31B model on coding tasks.
ollama run gemma4:31b-coding-mtp-bf16
What's Changed
- Update MLX and MLX-C with threading fixes by @dhiltgen in https://github.com/ollama/ollama/pull/15845
- go: bump to 1.26 by @ParthSareen in https://github.com/ollama/ollama/pull/15904
- Add Gemma 4 MTP speculative decoding by @pdevine in https://github.com/ollama/ollama/pull/15980
Full Changelog: https://github.com/ollama/ollama/compare/v0.23.0...v0.23.1-rc0

Claude Desktop
Claude Desktop is now supported with Ollama Launch.
Claude Cowork and Claude Code are supported within the Claude Desktop App.
ollama launch claude-desktop
Claude Cowork
Claude Code
Claude Code on the terminal can still be accessed through the CLI with:
ollama launch claude
Not supported yet
- Web Search (coming soon)
- Extensions
What's Changed
- Launch Claude Desktop with ollama launch claude
- The Ollama app now surfaces featured models from server-driven recommendations
- Fixed OpenClaw gateway timeout on Windows by enforcing IPv4 loopback (thanks @UniquePratham)
- Hardened Metal initialization to gracefully handle ggml kernel compilation failures
New Contributors
- @UniquePratham made their first contribution in https://github.com/ollama/ollama/pull/15726
Full Changelog: https://github.com/ollama/ollama/compare/v0.22.1...v0.23.0

What's Changed
- Updated the Gemma 4 renderer for thinking and tool calling improvements
- Model recommendations are now updated without updating Ollama
- Aligned the desktop app's launch page with ollama launch integrations
- Fixed the Poolside integration title in ollama launch
Full Changelog: https://github.com/ollama/ollama/compare/v0.22.0...v0.22.1

New models
- NVIDIA's Nemotron 3 Omni
- Poolside's first open-weight coding model - Laguna XS.2
Full Changelog: https://github.com/ollama/ollama/compare/v0.21.2...v0.22.0

What's Changed
- Improved reliability of the OpenClaw onboarding flow in ollama launch
- Recommended models in ollama launch now appear in a fixed, canonical order
- OpenClaw integration now bundles Ollama's web search plugin in OpenClaw
New Contributors
- @madflow made their first contribution in https://github.com/ollama/ollama/pull/15733
Full Changelog: https://github.com/ollama/ollama/compare/v0.21.1...v0.21.2

What's Changed
Kimi CLI
You can now install and run the Kimi CLI through Ollama.
ollama launch kimi --model kimi-k2.6:cloud
Kimi CLI with Kimi K2.6 excels at long horizon agentic execution tasks through a multi-agent system.
- MLX runner adds logprobs support for compatible models
- Faster MLX sampling with fused top-P and top-K in a single sort pass, plus repeat penalties applied in the sampler
- Improved MLX prompt tokenization by moving tokenization into request handler goroutines
- Better MLX thread safety for array management
- GLM4 MoE Lite performance improvement with a fused sigmoid router head
- Fixed model picker showing stale model after switching chats in the macOS app
- Fixed structured outputs for Gemma 4 when think=false
Full Changelog: https://github.com/ollama/ollama/compare/v0.21.0...v0.21.1

What's Changed
- launch: skip unchanged integration rewrite configration by @hoyyeva in #15491
- launch/openclaw: fix --yes flag behaviour to skip channels configuration by @hoyyeva in #15589
- launch: OpenCode inline config by @hoyyeva in #15586
- launch: add hermes by @ParthSareen in #15569
- launch: always list cloud recommendations first by @hoyyeva in #15593
- cmd/launch: add Copilot CLI integration by @scaryrawr in #15583
New Contributors
- @scaryrawr made their first contribution in #15583
Full Changelog: https://github.com/ollama/ollama/compare/v0.20.8-rc0...v0.21.0

What's Changed
- Fix quality of gemma:e2b and gemma:e4b when thinking is disabled
- ROCm: Update to ROCm 7.2.1 on Linux by @saman-amd in https://github.com/ollama/ollama/pull/15483
Full Changelog: https://github.com/ollama/ollama/compare/v0.20.6...v0.20.7

What's Changed
- Gemma 4 Tool Calling Improvements
- App bug fixes and improvements
- Parallel tool calling improvements
- Docs for Hermes Agent
Full Changelog: https://github.com/ollama/ollama/compare/v0.20.5...v0.20.6

OpenClaw channel setup with Ollama Launch
What's Changed
- OpenClaw channel setup: connect WhatsApp, Telegram, Discord, and other messaging channels through ollama launch openclaw
- Enable flash attention for Gemma 4 on compatible GPUs
- ollama launch openclaw now detects curl-based OpenCode installs at ~/.opencode/bin
- Fix /save command for models with safetensors-based architectures
New Contributors
- @sjhddh made their first contribution in https://github.com/ollama/ollama/pull/15424
Full Changelog: https://github.com/ollama/ollama/compare/v0.20.4...v0.20.5

What's Changed
- mlx: Improve M5 performance with NAX
- gemma4: enable flash attention
Full Changelog: https://github.com/ollama/ollama/compare/v0.20.3...v0.20.4

What's Changed
- Gemma 4 Tool Calling improvements
- Added latest models to Ollama App
- OpenClaw fixes for launching TUI
Full Changelog: https://github.com/ollama/ollama/compare/v0.20.2...v0.20.3

What's Changed
- app: default app home view to new chat instead of launch by @jmorganca in #15312
Full Changelog: v0.20.1...v0.20.2

Gemma 4
Effective 2B (E2B)
ollama run gemma4:e2b
Effective 4B (E4B)
ollama run gemma4:e4b
26B (Mixture of Experts model with 4B active parameters)
ollama run gemma4:26b
31B (Dense)
ollama run gemma4:31b
What's Changed
- docs: update pi docs by @ParthSareen in https://github.com/ollama/ollama/pull/15152
- mlx: respect tokenizer add_bos_token setting in pipeline by @dhiltgen in https://github.com/ollama/ollama/pull/15185
- tokenizer: add SentencePiece-style BPE support by @dhiltgen in https://github.com/ollama/ollama/pull/15162
Full Changelog: https://github.com/ollama/ollama/compare/v0.19.0...v0.20.0-rc0

What's Changed
- Ollama's app will now no longer incorrectly show "model is out of date"
- ollama launch pi now includes web search plugin that uses Ollama's web search
- Improved KV cache hit rate when using the Anthropic-compatible API
- Fixed tool call parsing issue with Qwen3.5 where tool calls would be output in thinking
- MLX runner will now create periodic snapshots during prompt processing
- Fixed KV cache snapshot memory leak in MLX runner
- Fixed issue where flash attention would be incorrectly enabled for grok models
- Fixed qwen3-next:80b not loading in Ollama
New Contributors
- @amatas made their first contribution in https://github.com/ollama/ollama/pull/15022
Full Changelog: https://github.com/ollama/ollama/compare/v0.18.3...v0.19.0

Ollama launch VS Code
Ollama can now directly launch models in VS Code.
ollama launch vscode # or code
What's Changed
- docs: update claude code and openclaw for web search and non-interactive mode by @ParthSareen in https://github.com/ollama/ollama/pull/14922
- cmd/launch: skip redundant config writes when model unchanged by @hoyyeva in https://github.com/ollama/ollama/pull/14941
- mlxrunner: share KV cache across conversations with common prefixes by @jessegross in https://github.com/ollama/ollama/pull/14887
- app: fix desktop app stuck loading when OLLAMA_HOST is an unspecified bind address by @hoyyeva in https://github.com/ollama/ollama/pull/14885
- parsers: robust xml tool repair by @BruceMacD in https://github.com/ollama/ollama/pull/14961
- add ability to turn on debug request logging by @drifkin in https://github.com/ollama/ollama/pull/14106
- Fix mlxrunner subprocess deadlocks by @jessegross in https://github.com/ollama/ollama/pull/14919
- launch: skip openclaw gateway health check when no daemon install by @BruceMacD in https://github.com/ollama/ollama/pull/14984
- docs: nemoclaw integration by @BruceMacD in https://github.com/ollama/ollama/pull/14962
- mlx: update as of 3/23 by @dhiltgen in https://github.com/ollama/ollama/pull/14789
- docs: add Claude Code with Telegram guide by @ParthSareen in https://github.com/ollama/ollama/pull/15026
- mlxrunner: support partial match on pure transformer caches by @jessegross in https://github.com/ollama/ollama/pull/14985
- mlx: add mxfp4/mxfp8/nvfp4 importing by @pdevine in https://github.com/ollama/ollama/pull/15015
- integration: improve ability to test individual models by @dhiltgen in https://github.com/ollama/ollama/pull/14948
- ci: fix windows cgo compiler error by @dhiltgen in https://github.com/ollama/ollama/pull/15046
Full Changelog: https://github.com/ollama/ollama/compare/v0.18.2...v0.18.3

What's Changed
- Add extra check to ensure npm and git are installed before installing OpenClaw
- Claude Code will now be faster when run locally, due to preventing cache breakages
- Fix to correctly support ollama launch openclaw --model <model>
- Register Ollama's websearch package correctly for OpenClaw
Full Changelog: https://github.com/ollama/ollama/compare/v0.18.1...v0.18.2

Web Search and Fetch in OpenClaw
Ollama now ships with web search and web fetch plugin for OpenClaw. This allows Ollama's models (local or cloud) to search the web for the latest content and news. This also allows OpenClaw with Ollama to be able to fetch the web and extract readable content for processing. This feature does not execute JavaScript.
When using local models with web search in OpenClaw, ensure you are signed into Ollama with ollama signin
ollama launch openclaw
You can install web search directly into OpenClaw as a plugin if you already have OpenClaw configured and working:
Ollama web search plugin
openclaw plugins install @ollama/openclaw-web-search
Non-interactive (headless) mode for ollama launch
ollama launch can now be run in non-interactive mode.
Perfect for:
- Docker/containers: spin up an integration as a pipeline step to run evals, test prompts, or validate model behavior as part of your build. Tear it down when the job ends.
- CI/CD: Generate code reviews, security checks, and other tasks within your CI
- Scripts/automation: Kick off automated tasks with Ollama and claude code
- --model must be specified to run in headless mode
- --yes flag will auto-pull the model and skip any selectors
Try with: ollama launch claude --model kimi-k2.5:cloud --yes -- -p "how does this repository work?"
Use non-interactive mode in OpenClaw
You can ask your OpenClaw to run tasks using claude with subagents:
run ollama launch claude --model kimi-k2.5:cloud --yes -- -p "how does this repository work?" using a subagent
What's Changed
- rocm: doc driver constraints by @dhiltgen in https://github.com/ollama/ollama/pull/14833
- config: use native OpenClaw Ollama onboarding by @BruceMacD in https://github.com/ollama/ollama/pull/14829
- cmd/launch: add guards for headless mode by @ParthSareen in https://github.com/ollama/ollama/pull/14837
- bench: improve benchmarking tool by @dhiltgen in https://github.com/ollama/ollama/pull/14240
- cmd/launch: skip --install-daemon when systemd is unavailable by @BruceMacD in https://github.com/ollama/ollama/pull/14883
Full Changelog: https://github.com/ollama/ollama/compare/v0.18.0...v0.18.1

What's Changed
- cmd: TUI model ordering by @ParthSareen in https://github.com/ollama/ollama/pull/14814
- docs: document reasoning_effort support in OpenAI-compatible API by @flipbit03 in https://github.com/ollama/ollama/pull/14821
- launch: fix headless --yes integration flow and policy scoping by @ParthSareen in https://github.com/ollama/ollama/pull/14815
- api/create: always propagate :cloud source for cloud models by @drifkin in https://github.com/ollama/ollama/pull/14822
- launch: add compact window for claude code by @ParthSareen in https://github.com/ollama/ollama/pull/14823
- cmd: still populate ollama ls when using ollama run <model:cloud> by @drifkin in https://github.com/ollama/ollama/pull/14824
New Contributors
- @flipbit03 made their first contribution in https://github.com/ollama/ollama/pull/14821
Full Changelog: https://github.com/ollama/ollama/compare/v0.17.8-rc4...v0.18.0-rc0

What's Changed
- Allow thinking levels such as "medium" to correctly interpreted in Ollama's API for all thinking models
- Add context length to support compaction when using ollama launch
Full Changelog: https://github.com/ollama/ollama/compare/v0.17.6...v0.17.7

What's Changed
- Fixed issue where GLM-OCR would not work due to incorrect prompt rendering
- Fixed tool calling parsing and rendering for Qwen 3.5 models
New Contributors
- @Victor-Quqi made their first contribution in https://github.com/ollama/ollama/pull/14584
Full Changelog: https://github.com/ollama/ollama/compare/v0.17.5...v0.17.6

What's Changed
- Fixed crash in Qwen 3.5 models when split over GPU & CPU
- Fixed issue where Qwen 3.5 models would repeat themselves due to no presence penalty (note: you may have to redownload the qwen3.5 models: ollama pull qwen3.5:35b for example)
- ollama run --verbose will now show peak memory usage when using Ollama's MLX engine
- Fixed memory issues and crashes in MLX runner
- Fixed issue where Ollama would not be able to run models imported from Qwen3.5 GGUF files
Full Changelog: https://github.com/ollama/ollama/compare/v0.17.4...v0.17.5

What's Changed
- Tool call indices will now be included in parallel tool calls
Full Changelog: https://github.com/ollama/ollama/compare/v0.17.3...v0.17.4

What's Changed
- fix: window app crash on startup when update is pending by @hoyyeva in https://github.com/ollama/ollama/pull/14451
Full Changelog: https://github.com/ollama/ollama/compare/v0.17.1...v0.17.2

What's Changed
- models: add nemotron architecture support by @jmorganca in https://github.com/ollama/ollama/pull/14356
- Avoid Excessive MLX Memory Usage by @jessegross in https://github.com/ollama/ollama/pull/14341
- ui: use capability-based detection for web search by @hoyyeva in https://github.com/ollama/ollama/pull/14336
- mlxrunner: Fix duplicate log prefixes and reduce log noise by @jessegross in https://github.com/ollama/ollama/pull/14379
- model: improvements to LFM2 architectures by @jmorganca in https://github.com/ollama/ollama/pull/14368
- mlx: don't default to affine quantization for unquantized models by @jessegross in https://github.com/ollama/ollama/pull/14376
- app: add upgrade configuration to settings page by @hoyyeva in https://github.com/ollama/ollama/pull/13512
- update mlx-c bindings to 0.5.0 by @dhiltgen in https://github.com/ollama/ollama/pull/14380
Full Changelog: https://github.com/ollama/ollama/compare/v0.17.0...v0.17.1

What's Changed
- Improved tokenizer performance
- Ollama's macOS and Windows apps will now default to a context length based on available VRAM
New Contributors
- @natl-set made their first contribution in https://github.com/ollama/ollama/pull/14322
Full Changelog: https://github.com/ollama/ollama/compare/v0.16.3...v0.17.0

What's Changed
- New ollama launch cline added for the Cline CLI
- ollama launch <integration> will now always show the model picker
- Added Gemma 3, Llama and Qwen 3 architectures to MLX runner
New Contributors
- @hellosaumil made their first contribution in https://github.com/ollama/ollama/pull/14271
Full Changelog: https://github.com/ollama/ollama/compare/v0.16.2...v0.16.3

What's Changed
- ollama launch claude now supports searching the web when using :cloud models
- Fixed rendering issue when running ollama in PowerShell
- New setting in Ollama's app makes it easier to disable cloud models for sensitive and private tasks where data cannot leave your computer. For Linux or when running ollama serve manually, set OLLAMA_NO_CLOUD=1.
- Fixed issue where experimental image generation models would not run in 0.16.0 and 0.16.1
Full Changelog: https://github.com/ollama/ollama/compare/v0.16.1...v0.16.2-rc0

What's Changed
- Installing Ollama via the curl install script on macOS will now only prompt for your password if its required
- Installing Ollama via the iem install script in Windows will now show progress
- Image generation models will now respect the OLLAMA_LOAD_TIMEOUT variable
Full Changelog: https://github.com/ollama/ollama/compare/v0.16.0...v0.16.1

New models
- GLM-5: A strong reasoning and agentic model from Z.ai with 744B total parameters (40B active), built for complex systems engineering and long-horizon tasks.
- MiniMax-M2.5: a new state-of-the-art large language model designed for real-world productivity and coding tasks.
New ollama
The new ollama command makes it easy to launch your favorite apps with models using Ollama
Ollama screenshot 2026-02-12 at 04 48 55@2x
What's Changed
- Launch Pi with ollama launch pi
- Improvements to Ollama's MLX runner to support GLM-4.7-Flash
- Ctrl+G will now allow for editing text prompts when running a model
Full Changelog: v0.15.6...v0.16.0

What's Changed
- Fixed context limits when running ollama launch droid
- ollama launch will now download missing models instead of erroring
- Fixed bug where ollama launch claude would cause context compaction when providing images
Full Changelog: https://github.com/ollama/ollama/compare/v0.15.5...v0.15.6

New models
- GLM-OCR: GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder–decoder architecture.
- Qwen3-Coder-Next: a coding-focused language model from Alibaba's Qwen team, optimized for agentic coding workflows and local development.
Improvements to ollama launch
- ollama launch can now be provided arguments, for example ollama launch claude -- --resume
- ollama launch will now work run subagents when using ollama launch claude
- Ollama will now set context limits for a set of models when using ollama launch opencode
What's Changed
- Sub-agent support for ollama launch for planning, deep research, and similar tasks
- ollama signin will now open a browser window to make signing in easier
- Ollama will now default to the following context lengths based on VRAM:
- < 24 GiB VRAM: 4,096 context
- 24-48 GiB VRAM: 32,768 context
- >= 48 GiB VRAM: 262,144 context
- GLM-4.7-Flash support on Ollama's experimental MLX engine
- ollama signin will now open the browser to the connect page
- Fixed off by one error when using num_predict in the API
- Fixed issue where tokens from a previous sequence would be returned when hitting num_predict
New Contributors
- @avukmirovich made their first contribution in https://github.com/ollama/ollama/pull/13934
Full Changelog: https://github.com/ollama/ollama/compare/v0.15.4...v0.15.5

What's Changed
- ollama launch openclaw will now enter the standard OpenClaw onboarding flow if this has not yet been completed.
Full Changelog: https://github.com/ollama/ollama/compare/v0.15.3...v0.15.4

What's Changed
- New ollama launch clawdbot command for launching Clawdbot using Ollama models
Full Changelog: https://github.com/ollama/ollama/compare/v0.15.1...v0.15.2

New feature
ollama launch command to use Ollama's models with Claude Code, Codex, OpenCode, and Droid without separate configuration.
What's Changed
- x/imagegen: add image edit capabilities by @jmorganca in https://github.com/ollama/ollama/pull/13846
- cmd: handle Enter key pressed during model loading, render multiline better by @ParthSareen in https://github.com/ollama/ollama/pull/13839
- x/imagegen: replace memory estimation with actual weight size by @jmorganca in https://github.com/ollama/ollama/pull/13848
- cmd: ollama config command to help configure integrations to use Ollama by @ParthSareen in https://github.com/ollama/ollama/pull/13712
- x/imagegen: respect stream=false in /api/generate by @jmorganca in https://github.com/ollama/ollama/pull/13853
- cmd: ollama config fix droid model name configuration by @ParthSareen in https://github.com/ollama/ollama/pull/13856
- model: add MLA absorption for glm4moelite by @jmorganca in https://github.com/ollama/ollama/pull/13810
- x/imagegen: fix image editing support by @jmorganca in https://github.com/ollama/ollama/pull/13866
- Revert "model: add MLA absorption for glm4moelite" by @jmorganca in https://github.com/ollama/ollama/pull/13869
- Re-apply "model: add MLA absorption for glm4moelite" with fix by @jmorganca in https://github.com/ollama/ollama/pull/13870
- cmd: rename ollama config to ollama launch by @ParthSareen in https://github.com/ollama/ollama/pull/13871
- llama: fix fattn-tile shared memory overflow on sm_50/52 by @jmorganca in https://github.com/ollama/ollama/pull/13872
- llama: fix CUDA release build issues by @jmorganca in https://github.com/ollama/ollama/pull/13874
- Clean up the manifest and modelpath by @pdevine in https://github.com/ollama/ollama/pull/13807
- x/imagegen: remove qwen_image and qwen_image_edit models by @jmorganca in https://github.com/ollama/ollama/pull/13827

New models
- GLM-4.7-Flash: As the strongest model in the 30B class, GLM-4.7-Flash offers a new option for lightweight deployment that balances performance and efficiency.
- LFM2.5-1.2B-Thinking: LFM2.5 is a new family of hybrid models designed for on-device deployment.
What's Changed
- Fixed issue where Ollama's macOS app would interrupt system shutdown
- Fixed ollama create and ollama show commands for experimental models
- The /api/generate API can now be used for image generation
- Fixed minor issues in Nemotron-3-Nano tool parsing
- Fixed issue where removing an image generation model would cause it to first load
- Fixed issue where ollama rm would only stop the first model in the list if it were running
Full Changelog: https://github.com/ollama/ollama/compare/v0.14.2...v0.14.3-rc2

New models
- TranslateGemma: A new collection of open translation models built on Gemma 3, helping people communicate across 55 languages.
What's Changed
- Shift + Enter will now enter a newline in Ollama's CLI
- Improve /v1/responses API to better confirm to OpenResponses specification
New Contributors
- @yuhongsun96 made their first contribution in https://github.com/ollama/ollama/pull/13135
- @koaning made their first contribution in https://github.com/ollama/ollama/pull/13326
Full Changelog: https://github.com/ollama/ollama/compare/v0.14.1...v0.14.2

What's Changed
- fix macOS auto-update signature verification failure
New Contributors
- @joshxfi made their first contribution in https://github.com/ollama/ollama/pull/13711
- @maternion made their first contribution in https://github.com/ollama/ollama/pull/13709
Full Changelog: https://github.com/ollama/ollama/compare/v0.14.0...v0.14.1

What's Changed
- ollama run --experimental CLI will now open a new Ollama CLI that includes an agent loop and the bash tool
- Anthropic API compatibility: support for the /v1/messages API
- A new REQUIRES command for the Modelfile allows declaring which version of Ollama is required for the model
- For older models, Ollama will avoid an integer underflow on low VRAM systems during memory estimation
- More accurate VRAM measurements for AMD iGPUs
- Ollama's app will now highlight swift soure code
- An error will now return when embeddings return NaN or -Inf
- Ollama's Linux install bundles files now use zst compression
- New experimental support for image generation models, powered by MLX
New Contributors
- @Vallabh-1504 made their first contribution in https://github.com/ollama/ollama/pull/13550
- @majiayu000 made their first contribution in https://github.com/ollama/ollama/pull/13596
- @harrykiselev made their first contribution in https://github.com/ollama/ollama/pull/13615
Full Changelog: https://github.com/ollama/ollama/compare/v0.13.5...v0.14.0-rc2

New Model
- Google's FunctionGemma is now available on Ollama
What's Changed
- bert architecture models now run on Ollama's engine
- Added built-in renderer & tool parsing capabilities for DeepSeek-V3.1
- Fixed issue where nested properties in tools may not have been rendered properly
New Contributors
- @familom made their first contribution in https://github.com/ollama/ollama/pull/13220
- @nathannewyen made their first contribution in https://github.com/ollama/ollama/pull/13469
Full Changelog: https://github.com/ollama/ollama/compare/v0.13.4...v0.13.5-rc1

New Models
- Nemotron 3 Nano: A new Standard for Efficient, Open, and Intelligent Agentic Models
- Olmo 3 and Olmo 3.1: A series of Open language models designed to enable the science of language models. These models are pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets.
What's Changed
- Enable Ollama engine by default for remaining models
- Enable Flash Attention automatically for models by default
- Fixed handling of long contexts with Gemma 3 models
- Fixed issue that would occur with Gemma 3 QAT models or other models imported with the Gemma 3 architecture
New Contributors
- @familom made their first contribution in https://github.com/ollama/ollama/pull/13220
Full Changelog: https://github.com/ollama/ollama/compare/v0.13.3...v0.13.4-rc0

New models
- Devstral-Small-2: 24B model that excels at using tools to explore codebases, editing multiple files and power software engineering agents.
- rnj-1: Rnj-1 is a family of 8B parameter open-weight, dense models trained from scratch by Essential AI, optimized for code and STEM with capabilities on par with SOTA open-weight models.
- nomic-embed-text-v2: nomic-embed-text-v2-moe is a multilingual MoE text embedding model that excels at multilingual retrieval.
What's Changed
- Improved truncation logic when using /api/embed and /v1/embeddings
- Extend Gemma 3 architecture to support rnj-1 model
- Fix error that would occur when running qwen2.5vl with image input
Full Changelog: https://github.com/ollama/ollama/compare/v0.13.2...v0.13.3

What's Changed
- Flash attention is now enabled by default for vision models such as mistral-3, gemma3, qwen3-vl and more. This improves memory utilization and performance when providing images as input.
- Fixed GPU detection on multi-GPU CUDA machines
- Fixed issue where deepseek-v3.1 would always think even with thinking is disabled in Ollama's app
New Contributors
- @chengcheng84 made their first contribution in https://github.com/ollama/ollama/pull/13265
- @nathan-hook made their first contribution in https://github.com/ollama/ollama/pull/13256
Full Changelog: https://github.com/ollama/ollama/compare/v0.13.1...v0.13.2-rc0

New models
- Ministral-3: The Ministral 3 family is designed for edge deployment, capable of running on a wide range of hardware.
What's Changed
- nomic-embed-text will now use Ollama's engine by default
- Tool calling support for cogito-v2.1
- Fixed issues with CUDA VRAM discovery
- Fixed link to docs in Ollama's app
Full Changelog: https://github.com/ollama/ollama/compare/v0.13.0...v0.13.1-rc2

New models
- DeepSeek-OCR: DeepSeek-OCR uses optical 2D mapping to compress long contexts, achieving high OCR precision with reduced vision tokens and demonstrating practical value in document processing.
- Cogito-V2.1: instruction tuned generative models, currently the best open-weight LLM by a US company
DeepSeek-OCR
DeepSeek-OCR is now available on Ollama. Example inputs:
ollama run deepseek-ocr "/path/to/image\n<|grounding|>Given the layout of the image."
ollama run deepseek-ocr "/path/to/image\nFree OCR."
ollama run deepseek-ocr "/path/to/image\nParse the figure."
ollama run deepseek-ocr "/path/to/image\nExtract the text in the image."
ollama run deepseek-ocr "/path/to/image\n<|grounding|>Convert the document to markdown."
New bench tool
Ollama's GitHub repo now includes a bench tool that can be used to test model performance. For the time being this is a separate tool that can be built in the Ollama GitHub repository:
First, install Go. Then from the root of the Ollama repository run:
go run ./cmd/bench.go -model gpt-oss:20b
For more information see the tool's documentation
What's Changed
- DeepSeek-OCR is now supported
- DeepSeek-V3.1 architecture is now supported in Ollama's engine
- Fixed performance issues that arose in Ollama 0.12.11 on CUDA
- Fixed issue where Linux install packages were missing required Vulkan libraries
- Improved CPU and memory detection while in containers/cgroups
- Improved VRAM information detection for AMD GPUs
- Improved KV cache performance to no longer require defragmentation
New Contributors
- @lnicola made their first contribution in https://github.com/ollama/ollama/pull/13096
- @vignesh1507 made their first contribution in https://github.com/ollama/ollama/pull/13078
- @pierwill made their first contribution in https://github.com/ollama/ollama/pull/12995
- @jjuliano made their first contribution in https://github.com/ollama/ollama/pull/11877
- @omahs made their first contribution in https://github.com/ollama/ollama/pull/10683
- @SiLeader made their first contribution in https://github.com/ollama/ollama/pull/10292
- @ssam18 made their first contribution in https://github.com/ollama/ollama/pull/13124
- @seolyam made their first contribution in https://github.com/ollama/ollama/pull/13116
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.11...v0.13.0-rc0

What's Changed
- Ollama's API and the OpenAI-compatible API now supports Logprobs
- Ollama's new app now supports WebP images
- Improved rendering performance in Ollama's new app, especially when rendering code
- The "required" field in tool definitions will now be omitted if not specified
- Fixed issue where "tool_call_id" would be omitted when using the OpenAI-compatible API.
- Fixed issue where ollama create would import data from both consolidated.safetensors and other safetensor files.
- Ollama will now prefer dedicated GPUs over iGPUs when scheduling models
- Vulkan can now be enabled by setting OLLAMA_VULKAN=1. For example: OLLAMA_VULKAN=1 ollama serve
New Contributors
- @mags0ft made their first contribution in https://github.com/ollama/ollama/pull/11371
- @macarronesc made their first contribution in https://github.com/ollama/ollama/pull/12973
- @breatn made their first contribution in https://github.com/ollama/ollama/pull/12985
- @cybardev made their first contribution in https://github.com/ollama/ollama/pull/13045
- @baptistejamin made their first contribution in https://github.com/ollama/ollama/pull/12899
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.10...v0.12.11-rc1

ollama run now works with embedding models
ollama run can now run embedding models to generate vector embeddings from text:
ollama run embeddinggemma "Hello world"
Content can also be provided to ollama run via standard input:
echo "Hello world" | ollama run embeddinggemma
What's Changed
- Fixed errors when running qwen3-vl:235b and qwen3-vl:235b-instruct
- Enable flash attention for Vulkan (currently needs to be built from source)
- Add Vulkan memory detection for Intel GPU using DXGI+PDH
- Ollama will now return tool call IDs from the /api/chat API
- Fixed hanging due to CPU discovery
- Ollama will now show login instructions when switching to a cloud model in interactive mode
- Fix reading stale VRAM data
- ollama run now works with embedding models
New Contributors
- @ryanycoleman made their first contribution in https://github.com/ollama/ollama/pull/11740
- @Rajathbail made their first contribution in https://github.com/ollama/ollama/pull/12929
- @virajwad made their first contribution in https://github.com/ollama/ollama/pull/12664
- @AXYZdong made their first contribution in https://github.com/ollama/ollama/pull/8601
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.9...v0.12.10

What's Changed
- Fix performance regression on CPU only systems
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.8...v0.12.9-rc0

What's Changed
- qwen3-vl performance improvements, including flash attention support by default
- qwen3-vl will now output less leading whitespace in the response when thinking
- Fixed issue where deepseek-v3.1 thinking could not be disabled in Ollama's new app
- Fixed issue where qwen3-vl would fail to interpret images with transparent backgrounds
- Ollama will now stop running a model before removing it via ollama rm
- Fixed issue where prompt processing would be slower on Ollama's engine
- Ignore unsupported iGPUs when doing device discovery on Windows
New Contributors
- @athshh made their first contribution in https://github.com/ollama/ollama/pull/12822
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.7...v0.12.8

New models
- Qwen3-VL: Qwen3-VL is now available in all parameter sizes ranging from 2B to 235B
- MiniMax-M2: a 230 Billion parameter model built for coding & agentic workflows available on Ollama's cloud
Add files and adjust thinking levels in Ollama's new app
Ollama's new app now includes a way to add one or many files when prompting the model:
For better responses, thinking levels can now be adjusted for the gpt-oss models:
New API documentation
New API documentation is available for Ollama's API: https://docs.ollama.com/api
What's Changed
- Model load failures now include more information on Windows
- Fixed embedding results being incorrect when running embeddinggemma
- Fixed gemma3n on Vulkan backend
- Increased time allocated for ROCm to discover devices
- Fixed truncation error when generating embeddings
- Fixed request status code when running cloud models
- The OpenAI-compatible /v1/embeddings endpoint now supports encoding_format parameter
- Ollama will now parse tool calls that don't conform to {"name": name, "arguments": args} (thanks @rick-github!)
- Fixed prompt processing reporting in the llama runner
- Increase speed when scheduling models
- Fixed issue where FROM <model> would not inherit RENDERER or PARSER commands
New Contributors
- @npardal made their first contribution in https://github.com/ollama/ollama/pull/12715
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.6...v0.12.7-rc0

What's Changed
- Flash attention is now enabled by default for Gemma 3, improving performance and memory utilization
- Fixed issue where Ollama would hang while generating responses
- Fixed issue where qwen3-coder would act in raw mode when using /api/generate or ollama run qwen3-coder <prompt>
- Fixed qwen3-embedding providing invalid results
- Ollama will now evict models correctly when num_gpu is set
- Fixed issue where tool_index with a value of 0 would not be sent to the model
Experimental Vulkan Support
Experimental support for Vulkan is now available when you build locally from source. This will enable additional GPUs from AMD, and Intel which are not currently supported by Ollama. To build locally, install the Vulkan SDK and set VULKAN_SDK in your environment, then follow the developer instructions. In a future release, Vulkan support will be included in the binary release as well. Please file issues if you run into any problems.
New Contributors
- @yajianggroup made their first contribution in https://github.com/ollama/ollama/pull/12377
- @inforithmics made their first contribution in https://github.com/ollama/ollama/pull/11835
- @sbhavani made their first contribution in https://github.com/ollama/ollama/pull/12619
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.5...v0.12.6

What's Changed
- Ollama's app will now wait until Ollama is running to allow for a conversation to be started
- Fixed issue where "think": false would show an error instead of being silently ignored
- Fixed deepseek-r1 output issues
New Contributors
- @shengxinjing made their first contribution in https://github.com/ollama/ollama/pull/12415
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.4...v0.12.5-rc0

New models
- DeepSeek-V3.1-Terminus: DeepSeek-V3.1-Terminus is a hybrid model that supports both thinking mode and non-thinking mode. It delivers more stable & reliable outputs across benchmarks compared to the previous version:
Run on Ollama's cloud:
ollama run deepseek-v3.1:671b-cloud
Run locally (requires 500GB+ of VRAM)
ollama run deepseek-v3.1
- Kimi-K2-Instruct-0905: Kimi K2-Instruct-0905 is the latest, most capable version of Kimi K2. It is a state-of-the-art mixture-of-experts (MoE) language model, featuring 32 billion activated parameters and a total of 1 trillion parameters.
ollama run kimi-k2:1t-cloud
What's Changed
- Fixed issue where tool calls provided as stringified JSON would not be parsed correctly
- ollama push will now provide a URL to follow to sign in
- Fixed issues where qwen3-coder would output unicode characters incorrectly
- Fix issue where loading a model with /load would crash
New Contributors
- @gr4ceG made their first contribution in https://github.com/ollama/ollama/pull/12385
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.2...v0.12.3

What's Changed
- Models with Qwen3's architecture including MoE now run in Ollama's new engine
- Fixed issue where built-in tools for gpt-oss were not being rendered correctly
- Support multi-regex pretokenizers in Ollama's new engine
- Ollama's new engine can now load tensors by matching a prefix or suffix
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.1...v0.12.2

New models
- Qwen3 Embedding: state of the art open embedding model by the Qwen team
What's Changed
- Fixed issue where Gemma3 QAT models would not output correct tokens
- Fix issue where & characters in Qwen3-Coder would not be parsed correctly when function calling
- Fixed issues where ollama signin would not work properly on Linux
Full Changelog: https://github.com/ollama/ollama/compare/v0.12.0...v0.12.1

What's Changed
- Models with the Bert architecture now run on Ollama's engine
- Models with the Qwen 3 architecture now run on Ollama's engine
- Fix issue where older NVIDIA GPUs would not be detected if newer drivers were installed
- qwen3-coder now supports tools
- Fixed issue where models would not be imported correctly with ollama create
- Ollama will skip parsing the initial <think> if provided in the prompt for /api/generate by @rick-github
New Contributors
- @egyptianbman made their first contribution in https://github.com/ollama/ollama/pull/12300
- @russcoss made their first contribution in https://github.com/ollama/ollama/pull/12280
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.11...v0.12.0-rc0

What's Changed
- Support for CUDA 13
- Improved memory usage when using gpt-oss in Ollama's app
- Better scrolling better in Ollama's app when submitting long prompts
- Cmd +/- will now zoom and shrink text in Ollama's app
- Assistant messages can now by copied in Ollama's app
- Fixed error that would occur when attempting to import satefensor files by @rick-github in https://github.com/ollama/ollama/pull/12176
- Improved memory estimates for hybrid and recurrent models by @gabe-l-hart in https://github.com/ollama/ollama/pull/12186
- Fixed error that would occur when when batch size was greater than context length
- Flash attention & KV cache quantization validation fixes by @jessegross in https://github.com/ollama/ollama/pull/12231
- Add dimensions field to embed requests by @mxyng in https://github.com/ollama/ollama/pull/12242
- Enable new memory estimates in Ollama's new engine by default by @jessegross in https://github.com/ollama/ollama/pull/12252
- Ollama will no longer load split vision models in the Ollama engine by @jessegross in https://github.com/ollama/ollama/pull/12241
New Contributors
- @KashyapTan made their first contribution in https://github.com/ollama/ollama/pull/12188
- @carbonatedWaterOrg made their first contribution in https://github.com/ollama/ollama/pull/12230
- @fengyuchuanshen made their first contribution in https://github.com/ollama/ollama/pull/12249
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.10...v0.11.11-rc0

New models
- EmbeddingGemma a new open embedding model that delivers best-in-class performance for its size
What's Changed
- Support for EmbeddingGemma
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.9...v0.11.10

What's Changed
- gpt-oss now has flash attention enabled by default for systems that support it
- Improved load times for gpt-oss
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.7...v0.11.8

DeepSeek-V3.1
DeepSeek-V3.1 is now available to run via Ollama.
This model supports hybrid thinking, meaning thinking can be enabled or disabled by setting think in Ollama's API:
curl http://localhost:11434/api/chat -d '{
"model": "deepseek-v3.1",
"messages": [
{
"role": "user",
"content": "why is the sky blue?"
}
],
"think": true
}'
In Ollama's CLI, thinking can be enabled or disabled by running the /set think or /set nothink commands.
Turbo (in preview)
DeepSeek-V3.1 has over 671B parameters, and so a large amount of VRAM is required to run it. Ollama's Turbo mode (in preview) provides access to powerful hardware in the cloud you can use to run the model.
Turbo via Ollama's app
Screenshot 2025-08-25 at 1 23 37 PM
1. Download Ollama for macOS or Windows
2. Select deepseek-v3.1:671b from the model selector
3. Enable Turbo
Turbo via Ollama's CLI and libraries
1. Create an account on ollama.com/signup
2. Follow the docs for Ollama's CLI to upload authenticate your Ollama installation
3. Run the following:
OLLAMA_HOST=ollama.com ollama run deepseek-v3.1
For instructions on using Turbo with Ollama's Python and JavaScript library, see the docs
What's Changed
- Fixed issue where multiple models would not be loaded on CPU-only systems
- Ollama will now work with models who skip outputting the initial<think> tag (e.g. DeepSeek-V3.1)
- Fixed issue where text would be emitted when there is no opening <think> tag from a model
- Fixed issue where tool calls containing { or } would not be parsed correctly
New Contributors
- @zoupingshi made their first contribution in #12028
Full Changelog: v0.11.6...v0.11.7

What's Changed
- Ollama's app will now switch between chats faster
- Improved layout of messages in Ollama's app
- Fixed issue where command prompt would show when Ollama's app detected an old version of Ollama running
- Improved performance when using flash attention
- Fixed boundary case when encoding text using BPE
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.5...v0.11.6

What's Changed
- Performance improvements for the gpt-oss models
- Improved multi-GPU scheduling and reduced VRAM allocation when using more than 2 GPUs
- Ollama's new app will now remember default selections for default model, Turbo and Web Search between restarts
- Fix error when parsing bad harmony tool calls
- OLLAMA_FLASH_ATTENTION=1 will also enable flash attention for pure-CPU models
- Fixed OpenAI-compatible API not supporting reasoning_effort
- Reduced size of installation on Windows and Linux
- New Memory Management by @jessegross in https://github.com/ollama/ollama/pull/11090
New Contributors
- @vorburger made their first contribution in https://github.com/ollama/ollama/pull/11755
- @dan-and made their first contribution in https://github.com/ollama/ollama/pull/10678
- @youzichuan made their first contribution in https://github.com/ollama/ollama/pull/11880
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.4...v0.11.5-rc1

What's Changed
- openai: allow for content and tool calls in the same message by @drifkin in https://github.com/ollama/ollama/pull/11759
- openai: when converting role=tool messages, propagate the tool name by @drifkin in https://github.com/ollama/ollama/pull/11761
- openai: always provide reasoning by @drifkin in https://github.com/ollama/ollama/pull/11765
New Contributors
- @gao-feng made their first contribution in https://github.com/ollama/ollama/pull/11170
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.3...v0.11.4

What's Changed
- Fixed issue where gpt-oss would consume too much VRAM when split across GPU & CPU or multiple GPUs
- Statically link C++ libraries on windows for better compatibility
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.2...v0.11.3

What's Changed
- Fix crash in gpt-oss when using kv cache quanitization
- Fix oss-gpt bug with "currentDate" not defined
Full Changelog: https://github.com/ollama/ollama/compare/v0.11.1...v0.11.2

Welcome OpenAI's gpt-oss models
Ollama partners with OpenAI to bring its latest state-of-the-art open weight models to Ollama. The two models, 20B and 120B, bring a whole new local chat experience, and are designed for powerful reasoning, agentic tasks, and versatile developer use cases.
Feature highlights
- Agentic capabilities: Use the models’ native capabilities for function calling, web browsing (Ollama is providing a built-in web search that can be optionally enabled to augment the model with the latest information), python tool calls, and structured outputs.
- Full chain-of-thought: Gain complete access to the model's reasoning process, facilitating easier debugging and increased trust in outputs.
- Configurable reasoning effort: Easily adjust the reasoning effort (low, medium, high) based on your specific use case and latency needs.
- Fine-tunable: Fully customize models to your specific use case through parameter fine-tuning.
- Permissive Apache 2.0 license: Build freely without copyleft restrictions or patent risk—ideal for experimentation, customization, and commercial deployment.
Quantization - MXFP4 format
OpenAI utilizes quantization to reduce the memory footprint of the gpt-oss models. The models are post-trained with quantization of the mixture-of-experts (MoE) weights to MXFP4 format, where the weights are quantized to 4.25 bits per parameter. The MoE weights are responsible for 90+% of the total parameter count, and quantizing these to MXFP4 enables the smaller model to run on systems with as little as 16GB memory, and the larger model to fit on a single 80GB GPU.
Ollama is supporting the MXFP4 format natively without additional quantizations or conversions. New kernels are developed for Ollama’s new engine to support the MXFP4 format.
Ollama collaborated with OpenAI to benchmark against their reference implementations to ensure Ollama’s implementations have the same quality.
Get started
You can get started by downloading the latest Ollama version (v0.11)
The model can be downloaded directly in Ollama’s new app or via the terminal:
ollama run gpt-oss:20b
ollama run gpt-oss:120b
What's Changed
- kvcache: Enable SWA to retain additional entries by @jessegross in https://github.com/ollama/ollama/pull/11611
- kvcache: Log contents of cache when unable to find a slot by @jessegross in https://github.com/ollama/ollama/pull/11658
Full Changelog: https://github.com/ollama/ollama/compare/v0.10.1...v0.11.0

What's Changed
- Fixed unicode character input for Japanese and other languages in Ollama's new app
- Fixed AMD download URL in the logs for ollama serve
New Contributors
- @skools-here made their first contribution in https://github.com/ollama/ollama/pull/11579
Full Changelog: https://github.com/ollama/ollama/compare/v0.10.0...v0.10.1

What's Changed
- ollama ps will now show the context length of loaded models
- Improved performance in gemma3n models by 2-3x
- Parallel request processing now defaults to 1. For more details, see the FAQ
- Fixed issue where tool calling would not work correctly with granite3.3 and mistral-nemo models
- Fixed issue where Ollama's tool calling would not work correctly if a tool's name was part of of another one, such as add and get_address
- Improved performance when using multiple GPUs by 10-30%
- Ollama's OpenAI-compatible API will now support WebP images
- Fixed issue where ollama show would report an error
- ollama run will more gracefully display errors
New Contributors
- @sncix made their first contribution in https://github.com/ollama/ollama/pull/11189
- @mfornet made their first contribution in https://github.com/ollama/ollama/pull/11425
- @haiyuewa made their first contribution in https://github.com/ollama/ollama/pull/11427
- @warting made their first contribution in https://github.com/ollama/ollama/pull/11461
- @ycomiti made their first contribution in https://github.com/ollama/ollama/pull/11462
- @minxinyi made their first contribution in https://github.com/ollama/ollama/pull/11502
- @ruyut made their first contribution in https://github.com/ollama/ollama/pull/11528
Full Changelog: https://github.com/ollama/ollama/compare/v0.9.6...v0.10.0

What's Changed
- Fixed styling issue in launch screen
- tool_name can now be provided in messages with "role": "tool" using the /api/chat endpoint
New Contributors
- @vrampal made their first contribution in https://github.com/ollama/ollama/pull/9681
Full Changelog: https://github.com/ollama/ollama/compare/v0.9.5...v0.9.6-rc0

Updates to Ollama for macOS and Windows
A new version of Ollama's macOS and Windows applications are now available. New improvements to the apps will be introduced over the coming releases:
- Download for macOS
- Download for Windows
New features
Expose Ollama on the network
Ollama can now be exposed on the network, allowing others to access Ollama on other devices or even over the internet. This is useful for having Ollama running on a powerful Mac, PC or Linux computer while making it accessible to less powerful devices.
Model directory
The directory in which models are stored can now be modified! This allows models to be stored on external hard disks or alternative directories than the default.
Smaller footprint and faster starting on macOS
The macOS app is now a native application and starts much faster while requiring a much smaller installation.
Additional changes in 0.9.5
- Fixed issue where the ollama CLI would not be installed by Ollama on macOS on startup
- Fixed issue where files in ollama-darwin.tgz were not notarized
- Add NativeMind to Community Integrations by @xukecheng in https://github.com/ollama/ollama/pull/11242
New Contributors
- @xukecheng made their first contribution in https://github.com/ollama/ollama/pull/11242

Updates to Ollama for macOS and Windows
A new version of Ollama's macOS and Windows applications are now available. New improvements to the apps will be introduced over the coming releases:
- Download for macOS
- Download for Windows
New features
Expose Ollama on the network
Ollama can now be exposed on the network, allowing others to access Ollama on other devices or even over the internet. This is useful for having Ollama running on a powerful Mac, PC or Linux computer while making it accessible to less powerful devices.
Model directory
The directory in which models are stored can now be modified! This allows models to be stored on external hard disks or alternative directories than the default.
Smaller footprint and faster starting on macOS
The macOS app is now a native application and starts much faster while requiring a much smaller installation.
What's Changed
- Fixed issue where tool calls would not be parsed when they were called with an empty set of arguments
- Fixed issue when quantizing models with the Gemma 3n architecture
Full Changelog: https://github.com/ollama/ollama/compare/v0.9.3...v0.9.4-rc5

What's Changed
- Fixed issue where errors would not be properly reported on Apple Silicon Macs
- Ollama will now limit context length to what the model was trained against to avoid strange overflow behavior
New Contributors
- @Aj-Seven made their first contribution in https://github.com/ollama/ollama/pull/11169
Full Changelog: https://github.com/ollama/ollama/compare/v0.9.2...v0.9.3

What's Changed
- Fixed issue where tool calls without parameters would not be returned correctly
- Fixed does not support generate errors
- Fixed issue where some special tokens would not be tokenized properly for some model architectures
New Contributors
- @NGC13009 made their first contribution in https://github.com/ollama/ollama/pull/11080
Full Changelog: https://github.com/ollama/ollama/compare/v0.9.1...v0.9.2

What's Changed
- Error messages that previously showed POST predict will now be more informative
- Improved tool calling reliability for some models
- Fixed issue on Windows where ollama run would not start Ollama automatically
New Contributors
- @JasonHonKL made their first contribution in https://github.com/ollama/ollama/pull/10174
- @hwittenborn made their first contribution in https://github.com/ollama/ollama/pull/10998
- @krzysztofjeziorny made their first contribution in https://github.com/ollama/ollama/pull/10973
Full Changelog: https://github.com/ollama/ollama/compare/v0.9.0...v0.9.1

New models
- DeepSeek-R1-0528: A new version of DeepSeek-R1 with significantly improved depth of reasoning and inference capabilities.
ollama pull deepseek-r1
Thinking mode
Ollama now supports thinking mode.
To enable thinking mode in Ollama, first download models that support thinking:
ollama pull deepseek-r1
ollama pull qwen3
Thinking
When running a model that supports thinking, Ollama will now display the model's thoughts:
% ollama run deepseek-r1
>>> How many Rs are in strawberry
Thinking...
First, I need to understand what the question is asking. It's asking how many letters 'R' are present in the word "strawberry."

Next, I'll examine each letter in the word individually.

I'll start from the beginning and count every occurrence of the letter 'R.'

After reviewing all the letters, I determine that there are three instances where the letter 'R' appears in the word "strawberry."
...done thinking.

There are three **Rs** in the word **"strawberry"**.
In Ollama's API, a model's thinking is now returned as a separate thinking field for easy parsing:
{
"message": {
"role": "assistant",
"thinking": "First, I need to understand what the question is asking. It's asking how many letters 'R' are present in the word "strawberry...",
"content": "There are **3** instances of the letter **R** in the word **"strawberry."**"
}
}
Turning thinking on and off
In the API, thinking can be enabled by passing "think": true and disabled by passing "think": false
curl http://localhost:11434/api/chat -d '{
"model": "deepseek-r1",
"messages": [
{
"role": "user",
"content": "Why is the sky blue?"
},
],
"think": true
}'
In Ollama's CLI, use /set think and /set nothink to enable and disable thinking.
What's Changed
- Add thinking mode support to Ollama
Full Changelog: https://github.com/ollama/ollama/compare/v0.8.0...v0.9.0

What's Changed
- Ollama will now stream responses with tool calls blog post
- Logs will now include better memory estimate debug information when running models in Ollama's engine.
New Contributors
- @hellotunamayo made their first contribution in https://github.com/ollama/ollama/pull/10790
Full Changelog: https://github.com/ollama/ollama/compare/v0.7.1...v0.8.0

What's Changed
- Improved model memory management to allocate sufficient memory to prevent crashes when running multimodal models in certain situations
- Enhanced memory estimation for models to prevent unintended memory offloading
- ollama show will now show ... when data is truncated
- Fixed crash that would occur with qwen2.5vl
- Fixed crash on Nvidia's CUDA for llama3.2-vision
- Support for Alibaba's Qwen 3 and Qwen 2 architectures in Ollama's new multimodal engine
New Contributors
- @ronxldwilson made their first contribution in https://github.com/ollama/ollama/pull/10763
- @DarkCaster made their first contribution in https://github.com/ollama/ollama/pull/10779
Full Changelog: https://github.com/ollama/ollama/compare/v0.7.0...v0.7.1-rc0

Ollama now supports multimodal models via Ollama’s new engine, starting with new vision multimodal models:
- Meta Llama 4
- Google Gemma 3
- Qwen 2.5 VL
- Mistral Small 3.1
- and more vision models.
What's Changed
- Fixed issue where a blank terminal window would appear when runnings models on Windows
- Fixed error that would occur when running llama4 on NVIDIA GPUs
- Reduced log level of key not found message
- Ollama will now correct remove quotes from image paths when sending images as input with ollama run
- Improved performance of importing safetensors models via ollama create
- Improved prompt processing speeds of Qwen3 MoE on macOS
- Fixed issue where providing large JSON schemas in structured output requests would result in an error
- Ollama's API will now return code 405 instead of 404 for methods that are not allowed
- Fixed issue where ollama processes would continue to run after a model was unloaded
New Contributors
- @ashokgelal made their first contribution in https://github.com/ollama/ollama/pull/8668
- @Aharon-Bensadoun made their first contribution in https://github.com/ollama/ollama/pull/9719
- @HardCodeDev777 made their first contribution in https://github.com/ollama/ollama/pull/10664
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.8...v0.7.0

What's Changed
- Performance improvements for Qwen 3 MoE models (30b-a3b and 235b-a22b) on NVIDIA and AMD GPUs
- Fixed GGML_ASSERT(tensor->op == GGML_OP_UNARY) failed issue caused by conflicting installations
- Fixed a memory leak that occurred when providing images as input
- ollama show will now correctly label older vision models such as llava
- Reduced out of memory errors by improving worst-case memory estimations
- Fix issue that resulted in a context canceled error
New Contributors
- @AliAhmedNada made their first contribution in https://github.com/ollama/ollama/pull/10522
- @HarshNevse made their first contribution in https://github.com/ollama/ollama/pull/10465
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.7...v0.6.8

What's Changed
- Add support for the Llama 4 multimodal models
- Increased default context window to 4096 tokens
- Fixed issue where image paths would not be recognized with ~ when being provided to ollama run
- Improved output quality when using JSON mode in certain scenarios
New Contributors
- @greengrass821 made their first contribution in https://github.com/ollama/ollama/pull/10339
- @richardshiue made their first contribution in https://github.com/ollama/ollama/pull/10335
- @aduermael made their first contribution in https://github.com/ollama/ollama/pull/10386
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.6...v0.6.7-rc0

New models
- IBM Granite 3.3: 2B and 8B models with 128K context length that have been fine-tuned for improved reasoning and instruction-following capabilities.
- DeepCoder: a fully open-Source 14B coder model at O3-mini level, with a 1.5B version also available.
What's Changed
- New, faster model downloading: OLLAMA_EXPERIMENT=client2 ollama serve to try running Ollama with this new downloader with improved performance and reliability when running ollama pull. Please share feedback here!
- Fixed memory leak issues when running Gemma 3, Mistral Small 3.1 and other models on Ollama
- Improved performance of ollama create when importing models from Safetensors
- Ollama will now allow tool function parameters with either a single type or an array of types by @rozgo
- Fixed certain out of memory issues from not reserving enough memory at startup
- Fix nondeterministic model unload order by @IreGaddr
- Include the items and $defs fields to properly handle array types in the API by @sheffler
- OpenAI-Beta headers are now included in the CORS safelist by @drifkin
- Fixed issue where model tensor data would be corrupted when importing models from Safetensors
New Contributors
- @drifkin made their first contribution in https://github.com/ollama/ollama/pull/10169
- @rozgo made their first contribution in https://github.com/ollama/ollama/pull/9434
- @qwerty108109 made their first contribution in https://github.com/ollama/ollama/pull/10168
- @IreGaddr made their first contribution in https://github.com/ollama/ollama/pull/10185
- @sheffler made their first contribution in https://github.com/ollama/ollama/pull/10091
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.5...v0.6.6-rc2

New models
- Mistral Small 3.1: the best performing vision model in its weight class.
What's Changed
- Support for Mistral Small 3.1
- Improved model loading times for Gemma 3 on network-backed filesystems such as Google Cloud Storage FUSE
New Contributors
- @danhipke made their first contribution in https://github.com/ollama/ollama/pull/10133
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.4...v0.6.5

What's Changed
- /api/show will now include model capabilities such as vision
- Fixed certain out-of-memory errors that would occur with parallel requests with Gemma 3
- Gemma 3 will now properly understand and output certain multilingual characters
- Fixed context shifting issues with models using the DeepSeek architecture
- Fixed issues with 0.6.3 where Gemma 3's output quality would worsen after 512 or 1024 tokens
- Added AMD RDNA4 support on Linux
New Contributors
- @saman-amd made their first contribution in https://github.com/ollama/ollama/pull/9878
- @leandroBorgesFerreira made their first contribution in https://github.com/ollama/ollama/pull/10042
- @Abyss-c0re made their first contribution in https://github.com/ollama/ollama/pull/9955
- @uggrock made their first contribution in https://github.com/ollama/ollama/pull/9983
- @IsAurora6 made their first contribution in https://github.com/ollama/ollama/pull/10057
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.3...v0.6.4

What's Changed
- New sliding window attention optimizations for Gemma 3, improving inference speed and memory allocation for long context windows.
- Improved loading speed of Gemma 3
- ollama create will now return the name of unsupported architectures
- Fixed error talloc->buffer_id >= 0 when running a model
- Fixed (int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs error when running a model
- ollama create will now automatically select the right template when importing Gemma 3 from safetensors
- ollama show -v will now correctly render boolean values as true or false
New Contributors
- @rylativity made their first contribution in https://github.com/ollama/ollama/pull/9874
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.2...v0.6.3

What's Changed
- Multiple images are now supported in Gemma 3
- Fixed issue where running Gemma 3 would consume a large amount of system memory
- ollama create --quantize now works when converting Gemma 3 from safetensors
- Fixed issue where /save would not work if running a model with / in the name
- Add support for AMD Strix Halo GPUs
New Contributors
- @zeozeozeo made their first contribution in https://github.com/ollama/ollama/pull/9800
- @louis030195 made their first contribution in https://github.com/ollama/ollama/pull/9786
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.1...v0.6.2-rc0

What's Changed
- Fixed issues where gemma3 would crash with "out of memory (OOM) errors" by improving memory estimation
- Loading an invalid model with /load will no longer exit ollama run
- Added Ctrl+P and Ctrl+N hotkeys for ollama run
- New ollama show -v or ollama show --verbose that will print additional model data
- Improved sampling parameters such as temperature and top_k to behave similar to other implementations
- New --verbose flag for ollama show that prints additional model data
New Contributors
- @Shane-XB-Qian made their first contribution in https://github.com/ollama/ollama/pull/9136
- @13rac1 made their first contribution in https://github.com/ollama/ollama/pull/9643
Full Changelog: https://github.com/ollama/ollama/compare/v0.6.0...v0.6.1

What's Changed
- Fixed error that would occur when running snowflake-arctic-embed and snowflake-arctic-embed2 models
New Contributors
- @iBreaker made their first contribution in https://github.com/ollama/ollama/pull/9565
- @zhu-xiaowei made their first contribution in https://github.com/ollama/ollama/pull/9540
- @dwt made their first contribution in https://github.com/ollama/ollama/pull/9569
- @vincentkoc made their first contribution in https://github.com/ollama/ollama/pull/9626
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.13...v0.6.0-rc0

New models
- Phi-4-Mini: Phi-4-mini brings significant enhancements in multilingual support, reasoning, and mathematics, and now, the long-awaited function calling feature is finally supported.
- Granite-3.2-Vision: A compact and efficient vision-language model, specifically designed for visual document understanding, enabling automated content extraction from tables, charts, infographics, plots, diagrams, and more.
- Command R7B Arabic: A new state-of-the-art version of the lightweight Command R7B model that excels in advanced Arabic language capabilities for enterprises in the Middle East and Northern Africa.
What's Changed
- The default context length can now be set with a new OLLAMA_CONTEXT_LENGTH environment variable. For example, to set the default context length to 8K, use: OLLAMA_CONTEXT_LENGTH=8192 ollama serve
- Ollama is now compiled for NVIDIA Blackwell
- Fixed issue where bf16 GGUF files could not be imported
- Ollama is now be able to accept requests from Visual Studio Code and Cursor by accepting requests from origins beginning with vscode-file://
New Contributors
- @prusnak made their first contribution in https://github.com/ollama/ollama/pull/8567
- @ChuanhuiLiu made their first contribution in https://github.com/ollama/ollama/pull/9346
- @Pekkari made their first contribution in https://github.com/ollama/ollama/pull/8211
- @eriestrisnadi made their first contribution in https://github.com/ollama/ollama/pull/9313
- @stevenh made their first contribution in https://github.com/ollama/ollama/pull/9360
- @wanghe-fit2cloud made their first contribution in https://github.com/ollama/ollama/pull/9312
- @Soulter made their first contribution in https://github.com/ollama/ollama/pull/9442
- @ibrahimcetin made their first contribution in https://github.com/ollama/ollama/pull/9411
- @remarkablemark made their first contribution in https://github.com/ollama/ollama/pull/9456
- @googs1025 made their first contribution in https://github.com/ollama/ollama/pull/9458
- @sunshine0523 made their first contribution in https://github.com/ollama/ollama/pull/9468
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.12...v0.5.13

New models
- Perplexity R1 1776: A version of the DeepSeek-R1 model that has been post trained to remove its refusal to respond to some sensitive topics.
What's Changed
- The OpenAI-compatible API will now return tool_calls if the model called a tool
- Performance on certain Intel Xeon processors should now be restored
- Fixed permission denied issues after installing Ollama on Linux
- Fixed issue where additional CPU libraries were included in the arm64 Linux install
- The progress bar will no longer flicker when running ollama pull
- Fixed issue where running a model would fail on Linux if Ollama was installed in a path with UTF-8 characters
- X-Stainless-Timeout will now be accepted as a header in the OpenAI API endpoints
New Contributors
- @buiducnhat made their first contribution in https://github.com/ollama/ollama/pull/9075
- @James-William-Kincaid-III made their first contribution in https://github.com/ollama/ollama/pull/9098
- @innightwolfsleep made their first contribution in https://github.com/ollama/ollama/pull/9150
- @ismdeep made their first contribution in https://github.com/ollama/ollama/pull/9108
- @benhaotang made their first contribution in https://github.com/ollama/ollama/pull/9138
- @jeremyschlatter made their first contribution in https://github.com/ollama/ollama/pull/9079
- @maninhill made their first contribution in https://github.com/ollama/ollama/pull/9212
- @xuzeyu91 made their first contribution in https://github.com/ollama/ollama/pull/9214
- @yuiseki made their first contribution in https://github.com/ollama/ollama/pull/9146
- @lucasthahn made their first contribution in https://github.com/ollama/ollama/pull/9237
- @danielekp made their first contribution in https://github.com/ollama/ollama/pull/9251
- @RockChinQ made their first contribution in https://github.com/ollama/ollama/pull/9272
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.11...v0.5.12

What's Changed
- Fixed The system cannot find the path specified errors when running models in some cases on Windows
- Fixed issue where running ollama serve on Intel Macs would not use CPU acceleration
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.10...v0.5.11

What's Changed
- Fixed issue on multi-GPU Windows and Linux machines where memory estimations would be incorrect
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.9...v0.5.10

New models
- DeepScaleR: A fine-tuned version of Deepseek-R1-Distilled-Qwen-1.5B that surpasses the performance of OpenAI’s o1-preview with just 1.5B parameters on popular math evaluations.
- OpenThinker: A fully open-source family of reasoning models built using a dataset derived by distilling DeepSeek-R1.
What's Changed
- Fixed llama runner process has terminated that would occur on Windows due to permissions issues
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.8...v0.5.9

What's Changed
- Fixed issue where using two FROM commands in Modelfile
- Support importing Command R and Command R+ architectures from safetensors
New Contributors
- @Gloryjaw made their first contribution in https://github.com/ollama/ollama/pull/8438
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.6...v0.5.7

What's Changed
- Fixed errors that would occur when running ollama create on Windows and when using absolute paths
New Contributors
- @steveberdy made their first contribution in https://github.com/ollama/ollama/pull/8352
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.5...v0.5.6

New models
- Phi-4: Phi 4 is a 14B parameter, state-of-the-art open model from Microsoft.
- Command R7B: the smallest model in Cohere's R series delivers top-tier speed, efficiency, and quality to build powerful AI applications on commodity GPUs and edge devices.
- DeepSeek-V3: A strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token.
- OLMo 2: a new family of 7B and 13B models trained on up to 5T tokens. These models are on par with or better than equivalently sized fully open models, and competitive with open-weight models such as Llama 3.1 on English academic benchmarks.
- Dolphin 3: the next generation of the Dolphin series of instruct-tuned models designed to be the ultimate general purpose local model, enabling coding, math, agentic, function calling, and general use cases.
- SmallThinker: A new small reasoning model fine-tuned from the Qwen 2.5 3B Instruct model.
- Granite 3.1 Dense: 2B and 8B text-only dense LLMs trained on over 12 trillion tokens of data, demonstrated significant improvements over their predecessors in performance and speed in IBM’s initial testing.
- Granite 3.1 MoE: 1B and 3B long-context mixture of experts (MoE) Granite models from IBM designed for low latency usage.
What's Changed
- The /api/create API endpoint that powers ollama create has been changed to improve conversion time and also accept a JSON object. Note: this change is not backwards compatible. If importing models, make sure you're using version 0.5.5 or later for both Ollama and the ollama CLI when running ollama create. If using ollama.create in the Python or JavaScript libraries, make sure to update to the latest version.
- Fixed runtime error that would occur when filling the model's context window
- Fixed crash that would occur when quotes were used in /save
- Fixed errors that would occur when sending x-stainless headers from OpenAI clients
New Contributors
- @Squishedmac made their first contribution in https://github.com/ollama/ollama/pull/8172
- @erusev made their first contribution in https://github.com/ollama/ollama/pull/7950
- @olumolu made their first contribution in https://github.com/ollama/ollama/pull/8227
- @paradoxical-dev made their first contribution in https://github.com/ollama/ollama/pull/8242
- @belfie13 made their first contribution in https://github.com/ollama/ollama/pull/8215
- @Docteur-RS made their first contribution in https://github.com/ollama/ollama/pull/7259
- @anxkhn made their first contribution in https://github.com/ollama/ollama/pull/8082
- @ubaldus made their first contribution in https://github.com/ollama/ollama/pull/8307
- @isamu made their first contribution in https://github.com/ollama/ollama/pull/8343
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.4...v0.5.5

New models
- Falcon3: A family of efficient AI models under 10B parameters performant in science, math, and coding through innovative training techniques.
What's Changed
- Fixed issue where providing null to format would result in an error
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.3...v0.5.4

What's Changed
- Fixed runtime errors on older Intel Macs
- Fixed issue where setting the format field to "" would cause an error
New Contributors
- @Askir made their first contribution in https://github.com/ollama/ollama/pull/8028
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.2...v0.5.3

New models
- EXAONE 3.5: a collection of instruction-tuned bilingual (English and Korean) generative models ranging from 2.4B to 32B parameters, developed and released by LG AI Research.
What's Changed
- Fixed issue where whitespace would get trimmed from prompt when images were provided
- Improved memory estimation when scheduling models
- OLLAMA_ORIGINS will now check hosts in a case insensitive manner
Note: the Linux ollama-linux-amd64.tgz directory structure has changed – if you manually install Ollama on Linux, make sure to retain the new directory layout and contents of the tar file.
New Contributors
- @yannickgloster made their first contribution in https://github.com/ollama/ollama/pull/7960
- @stweil made their first contribution in https://github.com/ollama/ollama/pull/7021
- @AidfulAI made their first contribution in https://github.com/ollama/ollama/pull/8024
- @taozuhong made their first contribution in https://github.com/ollama/ollama/pull/7948
- @philffm made their first contribution in https://github.com/ollama/ollama/pull/7202
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.1...v0.5.2

What's Changed
- Fixed issue where Ollama's API would generate JSON output when specifying "format": null
- Fixed issue where passing --format json to ollama run would cause an error
Full Changelog: https://github.com/ollama/ollama/compare/v0.5.0...v0.5.1

Structured outputs
Ollama now supports structured outputs, making it possible to constrain a model's output to a specific format defined by a JSON schema. The Ollama Python and JavaScript libraries have been updated to support structured outputs, together with Ollama's OpenAI-compatible API endpoints.
REST API
To use structured outputs in Ollama's generate or chat APIs, provide a JSON schema object in the format parameter:
curl -X POST http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
"model": "llama3.1",
"messages": [{"role": "user", "content": "Tell me about Canada."}],
"stream": false,
"format": {
"type": "object",
"properties": {
"name": {
"type": "string"
},
"capital": {
"type": "string"
},
"languages": {
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"name",
"capital",
"languages"
]
}
}'
Python library
Using the Ollama Python library, pass in the schema as a JSON object to the format parameter as either dict or use Pydantic (recommended) to serialize the schema using model_json_schema().
from ollama import chat
from pydantic import BaseModel

class Country(BaseModel):
name: str
capital: str
languages: list[str]

response = chat(
messages=[
{
'role': 'user',
'content': 'Tell me about Canada.',
}
],
model='llama3.1',
format=Country.model_json_schema(),
)

country = Country.model_validate_json(response.message.content)
print(country)
JavaScript library
Using the Ollama JavaScript library, pass in the schema as a JSON object to the format parameter as either object or use Zod (recommended) to serialize the schema using zodToJsonSchema():
import ollama from 'ollama';
import { z } from 'zod';
import { zodToJsonSchema } from 'zod-to-json-schema';

const Country = z.object({
name: z.string(),
capital: z.string(),
languages: z.array(z.string()),
});

const response = await ollama.chat({
model: 'llama3.1',
messages: [{ role: 'user', content: 'Tell me about Canada.' }],
format: zodToJsonSchema(Country),
});

const country = Country.parse(JSON.parse(response.message.content));
console.log(country);
What's Changed
- Fixed error importing model vocabulary files
- Experimental: new flag to set KV cache quantization to 4-bit (q4_0), 8-bit (q8_0) or 16-bit (f16). This reduces VRAM requirements for longer context windows.
- To enable for all models, use OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q4_0 ollama serve
- Note: in the future flash attention will be enabled by default where available, with kv cache quantization available on a per-model basis
- Thank you @sammcj for the contribution in in https://github.com/ollama/ollama/pull/7926
New Contributors
- @dmayboroda made their first contribution in https://github.com/ollama/ollama/pull/7906
- @Geometrein made their first contribution in https://github.com/ollama/ollama/pull/7908
- @owboson made their first contribution in https://github.com/ollama/ollama/pull/7693
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.7...v0.5.0

What's Changed
- The OpenAI compatible chat endpoint now includes index fields for tool calls, fixing issue when using chat.completions.stream and other methods
New Contributors
- @tcsenpai made their first contribution in https://github.com/ollama/ollama/pull/7216
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.6...v0.4.7

New models
- QwQ: an experimental research model by the Qwen team, focused on advancing AI reasoning capabilities.
What's Changed
- Tool calls will now be included in streaming responses
- Ollama will now provide an error when submitting SVG images
- Image tokens will now longer be counted in token counts when running a text-only model
New Contributors
- @ItzCrazyKns made their first contribution in https://github.com/ollama/ollama/pull/7667
- @ParthSareen made their first contribution in https://github.com/ollama/ollama/pull/7836
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.5...v0.4.6

What's Changed
- Fixed issue where HTTPS_PROXY and HTTP_PROXY environment variables would have no effect
- Ollama will now accept X-Stainless-Retry-Count used by many OpenAI API clients
- Fix issue where importing certain GGUF files would result in the incorrect quantization level
- ollama push will now print the uploaded model URL on ollama.com
New Contributors
- @rrg92 made their first contribution in https://github.com/ollama/ollama/pull/7438
- @oza6ut0ne made their first contribution in https://github.com/ollama/ollama/pull/6910
- @josStorer made their first contribution in https://github.com/ollama/ollama/pull/4118
- @patcher9 made their first contribution in https://github.com/ollama/ollama/pull/7811
- @erikos made their first contribution in https://github.com/ollama/ollama/pull/3591
- @sunnybak made their first contribution in https://github.com/ollama/ollama/pull/7831
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.4...v0.4.5

New models
- Marco-o1: An open large reasoning model for real-world solutions by the Alibaba International Digital Commerce Group (AIDC-AI).
What's Changed
- Fixed issue where Ollama would freeze when processing requests in parallel (e.g. when using code completion tools)
- Redirecting output to a file no longer outputs progress bars or spinners
New Contributors
- @zhanluxianshen made their first contribution in https://github.com/ollama/ollama/pull/7308
- @EliasPereirah made their first contribution in https://github.com/ollama/ollama/pull/7084
- @edwinjhlee made their first contribution in https://github.com/ollama/ollama/pull/5191
- @h1ddenpr0cess20 made their first contribution in https://github.com/ollama/ollama/pull/7325
- @mikelolasagasti made their first contribution in https://github.com/ollama/ollama/pull/7310
- @Leon-Sander made their first contribution in https://github.com/ollama/ollama/pull/6931
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.3...v0.4.4

New Models
- Tülu 3: Tülu 3 is a leading instruction following model family, offering fully open-source data, code, and recipes by the The Allen Institute for AI.
- Mistral Large: a new version of Mistral Large with improved Long Context, Function Calling and System Prompt support.
What's Changed
- Fixed crash that occurred on macOS devices with low memory
- Improved performance issues that occurred in Ollama versions 0.4.0-0.4.2
- Fixed issue that would cause granite3-dense to generate empty responses
- Fixed crashes and hanging caused by KV cache management
New Contributors
- @vinhnx made their first contribution in https://github.com/ollama/ollama/pull/7706
- @tusharad made their first contribution in https://github.com/ollama/ollama/pull/7451
- @DariusKocar made their first contribution in https://github.com/ollama/ollama/pull/7431
- @nbonamy made their first contribution in https://github.com/ollama/ollama/pull/7713
- @zeitlings made their first contribution in https://github.com/ollama/ollama/pull/7724
- @gkamer8 made their first contribution in https://github.com/ollama/ollama/pull/7746
- @iamrohitanshu made their first contribution in https://github.com/ollama/ollama/pull/7764
- @adarshM84 made their first contribution in https://github.com/ollama/ollama/pull/7707
- @thewh1teagle made their first contribution in https://github.com/ollama/ollama/pull/7607
- @marcusziade made their first contribution in https://github.com/ollama/ollama/pull/7383
- @emirsahin1 made their first contribution in https://github.com/ollama/ollama/pull/5931
- @MagicFun1241 made their first contribution in https://github.com/ollama/ollama/pull/7705
- @anan1213095357 made their first contribution in https://github.com/ollama/ollama/pull/7221
- @drunkwcodes made their first contribution in https://github.com/ollama/ollama/pull/7064
- @andygill made their first contribution in https://github.com/ollama/ollama/pull/6945
- @eschnou made their first contribution in https://github.com/ollama/ollama/pull/5497
- @nicarq made their first contribution in https://github.com/ollama/ollama/pull/4877
- @chyok made their first contribution in https://github.com/ollama/ollama/pull/5412
- @elearningshow made their first contribution in https://github.com/ollama/ollama/pull/5575
- @aarushik93 made their first contribution in https://github.com/ollama/ollama/pull/6459
- @f-lombardo made their first contribution in https://github.com/ollama/ollama/pull/5679
- @dezoito made their first contribution in https://github.com/ollama/ollama/pull/4301
- @jakubburkiewicz made their first contribution in https://github.com/ollama/ollama/pull/4648
- @promptery made their first contribution in https://github.com/ollama/ollama/pull/7093
- @szczyglis-dev made their first contribution in https://github.com/ollama/ollama/pull/6503
- @k33g made their first contribution in https://github.com/ollama/ollama/pull/7191
- @tzolov made their first contribution in https://github.com/ollama/ollama/pull/5981
- @cbldev made their first contribution in https://github.com/ollama/ollama/pull/5381
- @emrgnt-cmplxty made their first contribution in https://github.com/ollama/ollama/pull/5587
- @xuyangbocn made their first contribution in https://github.com/ollama/ollama/pull/5633
- @lemonit-eric-mao made their first contribution in https://github.com/ollama/ollama/pull/7126
- @paulrobello made their first contribution in https://github.com/ollama/ollama/pull/5623
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.2...v0.4.3

What's Changed
- Make KV cache shifting more reliable
- Fix NVIDIA JetPack support
- Fix extraneous newlines when displaying template/system layers
- Fix llama3.2-vision on multiple CUDA GPUs
- Fix error when embeddings contains only whitespace
- Report line numbers for parser errors in create
New Contributors
- @arbusam made their first contribution in https://github.com/ollama/ollama/pull/7580
- @neomantra made their first contribution in https://github.com/ollama/ollama/pull/7604
- @frances720 made their first contribution in https://github.com/ollama/ollama/pull/7517
- @ivostoykov made their first contribution in https://github.com/ollama/ollama/pull/5827
- @prasad89 made their first contribution in https://github.com/ollama/ollama/pull/7521
- @joey5403 made their first contribution in https://github.com/ollama/ollama/pull/7418
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.1...v0.4.2-rc0

What's Changed
- Fixed an issue where the cuda v12 runner on linux was incorrectly linked
- Fixed an issue with multiple AMD GPUs on Windows resulting in gibberish output
- Fixed support for MacOS v11 on ARM systems
New Contributors
- @edmcman made their first contribution in https://github.com/ollama/ollama/pull/7527
Full Changelog: https://github.com/ollama/ollama/compare/v0.4.0...v0.4.1-rc0

Llama 3.2 Vision
Ollama 0.4 adds support for Llama 3.2 Vision. After upgrading or downloading Ollama, run:
ollama run llama3.2-vision
For the larger, 90B version of the model, run:
ollama run llama3.2-vision:90b
What's changed
- Support for Llama 3.2 Vision (i.e. Mllama) architecture
- Improved performance on new generation NVIDIA graphics cards (e.g. RTX 40 series)
- Sending follow on requests to vision models will now be much faster
- Fixed issues where stop sequences would not be detected correctly
- Ollama can now import models from Safetensors without a Modelfile when running ollama create my-model
- Fixed issue where redirecting output to a file on Windows would cause invalid characters to be written
- Fixed issue where invalid model data would cause Ollama to error
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.14...v0.4.0

New Models
- Granite 3 MoE: The IBM Granite 1B and 3B models are the first mixture of experts (MoE) Granite models from IBM designed for low latency usage.
- Granite 3 Dense: The IBM Granite 2B and 8B models are designed to support tool-based use cases and support for retrieval augmented generation (RAG), streamlining code generation, translation and bug fixing.
Thank you @gabe-l-hart for contributing Granite support to Ollama!
What's Changed
- Fix crashes for AMD GPUs with small system memory
- Fix error that would occur on macOS 11 Big Sur
- Fixed issue creating models from bf16 file types
- Improve CPU performance by improving default thread counts
- IBM granite/granitemoe architecture support by @gabe-l-hart in https://github.com/ollama/ollama/pull/6760
New Contributors
- @JHubi1 made their first contribution in https://github.com/ollama/ollama/pull/6465
- @gabe-l-hart made their first contribution in https://github.com/ollama/ollama/pull/6760
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.13...v0.3.14

New safety models
- Llama Guard 3: a series of models by Meta, fine-tuned for content safety classification of LLM inputs and responses.
- ShieldGemma: ShieldGemma is set of instruction tuned models from Google DeepMind for evaluating the safety of text prompt input and text output responses against a set of defined safety policies.
What's Changed
- Default model in quickstart is now llama3.2
- Fixed issue where ollama pull would leave connections when encountering an error
- ollama rm will now stop a model if it is running prior to deleting it
New Go runner (for contributors):
Ollama is transitioning to a new Go-based subprocess model runner with improved reliability and caching. This new runner found under the llama directory replaces the previous code under llm/ext_server. During this transition period, the new Go runner is opt-in at build time, and requires using a different approach to build. For more information see development.md.
New Contributors
- @zmldndx made their first contribution in https://github.com/ollama/ollama/pull/7027
- @hidden1nin made their first contribution in https://github.com/ollama/ollama/pull/7096
- @shifragoldstone made their first contribution in https://github.com/ollama/ollama/pull/7117
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.12...v0.3.13

New models
- Llama 3.2: Meta's Llama 3.2 goes small with 1B and 3B models.
- Qwen 2.5 Coder: The latest series of Code-Specific Qwen models, with significant improvements in code generation, code reasoning, and code fixing.
What's Changed
- Ollama now supports ARM Windows machines
- Fixed rare issue where Ollama would report a missing .dll file on Windows
- Fixed performance issue for Windows without GPUs
New Contributors
- @RyanMarten made their first contribution in https://github.com/ollama/ollama/pull/6841
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.11...v0.3.12

New models
- Solar-Pro-Preview: an advanced large language model (LLM) with 22 billion parameters designed to fit into a single GPU.
- Qwen 2.5: new multilingual Qwen models pretrained on Alibaba's latest large-scale dataset, encompassing up to 18 trillion tokens with support for a context window of up to 128K tokens.
- Bespoke-Minicheck: a state-of-the-art fact-checking model developed by Bespoke Labs.
- Mistral-Small: a lightweight 22B model designed for cost-effective use in tasks like translation and summarization.
- Reader-LM: A series of models that convert HTML content to Markdown content, which is useful for content conversion tasks.
What's Changed
- New ollama stop command to unload a running model
- Ollama will now show an error when importing a model with an invalid number of tokens in the vocabulary
- The ollama/ollama container image will now start running almost immediately, leading to 5s faster start times
- Fixed issue where ollama show would show excessive whitespace in the output
New Contributors
- @Palm1r made their first contribution in https://github.com/ollama/ollama/pull/6754
- @codefromthecrypt made their first contribution in https://github.com/ollama/ollama/pull/6744
- @ECuiDev made their first contribution in https://github.com/ollama/ollama/pull/6789
- @pepo-ec made their first contribution in https://github.com/ollama/ollama/pull/6818
- @yangfuhai made their first contribution in https://github.com/ollama/ollama/pull/6788
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.10...v0.3.11

New models
- MiniCPM-V: A powerful, GPT-4V level multi-modal model.
- Yi-Coder: a series of open-source code language models that delivers state-of-the-art coding performance with fewer than 10 billion parameters.
- DeepSeek-V2.5: An upgraded version of DeepSeek-V2 that integrates the general and coding abilities of both DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct.
What's Changed
- Fixed rare error that would occur for certain models when running ollama show
- CUDA 11 will now be used for older NVIDIA drivers that are not compatible with CUDA 12
- Fixed error when running ollama create to import Gemma 2 models from safetensors
- The OpenAI-compatible chat and completions APIs will no longer scale temperature and frequency_penalty
New Contributors
- @rayfiyo made their first contribution in https://github.com/ollama/ollama/pull/6577
- @vimalk78 made their first contribution in https://github.com/ollama/ollama/pull/6579
- @SnoopyTlion made their first contribution in https://github.com/ollama/ollama/pull/6587
- @jonathanhecl made their first contribution in https://github.com/ollama/ollama/pull/6583
- @presbrey made their first contribution in https://github.com/ollama/ollama/pull/6508
- @yeahdongcn made their first contribution in https://github.com/ollama/ollama/pull/5847
- @openvmp made their first contribution in https://github.com/ollama/ollama/pull/6605
- @FellowTraveler made their first contribution in https://github.com/ollama/ollama/pull/5664
- @mateuszmigas made their first contribution in https://github.com/ollama/ollama/pull/5514
- @kspviswa made their first contribution in https://github.com/ollama/ollama/pull/6624
- @erkinalp made their first contribution in https://github.com/ollama/ollama/pull/6450
- @Carter907 made their first contribution in https://github.com/ollama/ollama/pull/6391
- @mitar made their first contribution in https://github.com/ollama/ollama/pull/6421
- @kangfenmao made their first contribution in https://github.com/ollama/ollama/pull/6633
- @teilomillet made their first contribution in https://github.com/ollama/ollama/pull/6099
- @fujitatomoya made their first contribution in https://github.com/ollama/ollama/pull/6430
- @1runeberg made their first contribution in https://github.com/ollama/ollama/pull/6378
- @iscy made their first contribution in https://github.com/ollama/ollama/pull/6642
- @jake83741 made their first contribution in https://github.com/ollama/ollama/pull/6644
- @ArdaGnsrn made their first contribution in https://github.com/ollama/ollama/pull/6361
- @SilasMarvin made their first contribution in https://github.com/ollama/ollama/pull/5063
- @wangqingfree made their first contribution in https://github.com/ollama/ollama/pull/6035
- @vitaly-zdanevich made their first contribution in https://github.com/ollama/ollama/pull/5714
- @Tobix99 made their first contribution in https://github.com/ollama/ollama/pull/6656
- @bean5 made their first contribution in https://github.com/ollama/ollama/pull/5888
- @ZeyoYT made their first contribution in https://github.com/ollama/ollama/pull/4957
- @imoize made their first contribution in https://github.com/ollama/ollama/pull/6681
- @nickthecook made their first contribution in https://github.com/ollama/ollama/pull/6680
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.9...v0.3.10

What's Changed
- Fixed error that would occur when running Ollama on Linux machines with the ARM architecture
- Ollama will now show an improved error message when attempting to run unsupported models
- Fixed issue where Ollama would not auto-detect the chat template for Llama 3.1 models
- OLLAMA_HOST will now work with with URLs that contain paths
New Contributors
- @bryanhonof made their first contribution in https://github.com/ollama/ollama/pull/6074
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.8...v0.3.9

What's Changed
- Fixed error where the ollama CLI couldn't be found on the path when upgrading Ollama on Windows
New Contributors
- @seankhatiri made their first contribution in https://github.com/ollama/ollama/pull/6530
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.7...v0.3.8

New Models
- Hermes 3: Hermes 3 is the latest version of the flagship Hermes series of LLMs by Nous Research, which includes support for tool calling.
- Phi 3.5: A lightweight AI model with 3.8 billion parameters with performance overtaking similarly and larger sized models.
- SmolLM: A family of small models with 135M, 360M, and 1.7B parameters, trained on a new high-quality dataset.
What's Changed
- CUDA 12 support: improving performance by up to 10% on newer NVIDIA GPUs
- Improved performance of ollama pull and ollama push on slower connections
- Fixed issue where setting OLLAMA_NUM_PARALLEL would cause models to be reloaded on lower VRAM systems
- Ollama on Linux is now distributed as a tar.gz file, which contains the ollama binary along with required libraries.
New Contributors
- @pamelafox made their first contribution in https://github.com/ollama/ollama/pull/6345
- @eust-w made their first contribution in https://github.com/ollama/ollama/pull/5964
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.6...v0.3.7

What's Changed
- Fixed issue where /api/embed would return an error instead of loading the model when the input field was not provided.
- ollama create can now import Phi-3 models from Safetensors
- Added progress information to ollama create when importing GGUF files
- Ollama will now import GGUF files faster by minimizing file copies
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.5...v0.3.6

What's Changed
- Fixed Incorrect function error when downloading models on Windows
- Fixed issue where temporary files would not be cleaned up
- Fix rare error when Ollama would start up due to invalid model data
- Ollama will now provide an error instead of crashing on Windows when running models that are too large to fit into total memory
New Contributors
- @jessegross made their first contribution in https://github.com/ollama/ollama/pull/6145
- @rgbkrk made their first contribution in https://github.com/ollama/ollama/pull/5985
- @Nicholas42 made their first contribution in https://github.com/ollama/ollama/pull/6235
- @cognitivetech made their first contribution in https://github.com/ollama/ollama/pull/6305
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.4...v0.3.5

New embedding models
- BGE-M3: a large embedding model from BAAI distinguished for its versatility in Multi-Functionality, Multi-Linguality, and Multi-Granularity.
- BGE-Large: a large embedding model trained in english.
- Paraphrase-Multilingual: A multilingual embedding model trained on parallel data for 50+ languages.
New embedding endpoint with batch support
Ollama now supports a new API endpoint /api/embed for embedding generation:
curl http://localhost:11434/api/embed -d '{
"model": "all-minilm",
"input": ["Why is the sky blue?", "Why is the grass green?"]
}'
This API endpoint supports new features:
- Batches: generate embeddings for several documents in one request
- Normalized embeddings: embeddings are now normalized, improving similarity results
- Truncation: a new truncate parameter that will error if set to false
- Metrics: responses include load_duration, total_duration and prompt_eval_count metrics
See the API documentation for more details and examples.
What's Changed
- Fixed initial slow download speeds on Windows
- NUMA support will now be autodetected by Ollama to improve performance
- Fixed issue where the /api/embed would sometimes return embedding results out of order
New Contributors
- @av made their first contribution in https://github.com/ollama/ollama/pull/6147
- @sryu1 made their first contribution in https://github.com/ollama/ollama/pull/6151
- @rick-github made their first contribution in https://github.com/ollama/ollama/pull/6154
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.3...v0.3.4

What's Changed
- The /api/embed endpoint now returns statistics: total_duration, load_duration, and prompt_eval_count
- Added usage metrics to the /v1/embeddings OpenAI compatibility API
- Fixed issue where /api/generate would respond with an empty string if provided a context
- Fixed issue where /api/generate would return an incorrect value for context
- /show modefile will now render MESSAGE commands correctly
New Contributors
- @slouffka made their first contribution in https://github.com/ollama/ollama/pull/6115
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.2...v0.3.3

What's Changed
- Fixed issue where ollama pull would not resume download progress
- Fixed issue where phi3 would report an error on older versions
New Contributors
- @longseespace made their first contribution in https://github.com/ollama/ollama/pull/6096
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.1...v0.3.2

Google's Gemma 2 has a new 2B parameter small model!
ollama run gemma2:2b
New models
- Gemma 2 2B: A new 2B parameter model by Google DeepMind
What's Changed
- Added support for min_p sampling option
- ollama create will now autodetect required stop parameters when importing certain models
- Ollama on Windows will now show better error messages if required files are missing
- Fixed issue where /save would cause parameters to be saved incorrectly
- OpenAI-compatible API will now return a finish_reason of tool_calls if a tool call occured.
- Performance and reliability improvements when downloading models using ollama pull
- Ollama's Linux install script will now return a better error on unsupported CUDA versions
New Contributors
- @Robitx made their first contribution in https://github.com/ollama/ollama/pull/1825
- @hellerve made their first contribution in https://github.com/ollama/ollama/pull/6041
Full Changelog: https://github.com/ollama/ollama/compare/v0.3.0...v0.3.1

Tool support
Ollama now supports tool calling with popular models such as Llama 3.1. This enables a model to answer a given prompt using tool(s) it knows about, making it possible for models to perform more complex tasks or interact with the outside world.
Example tools include:
- Functions and APIs
- Web browsing
- Code interpreter
- much more!
https://github.com/user-attachments/assets/957ef0d6-e7b7-4168-8033-13b0fe5f7029
To use tools, provide the tools field when using Ollama's Chat API:
import ollama

response = ollama.chat(
model='llama3.1',
messages=[{'role': 'user', 'content': 'What is the weather in Toronto?'}],

# provide a weather checking tool to the model
tools=[{
'type': 'function',
'function': {
'name': 'get_current_weather',
'description': 'Get the current weather for a city',
'parameters': {
'type': 'object',
'properties': {
'city': {
'type': 'string',
'description': 'The name of the city',
},
},
'required': ['city'],
},
},
},
],
)

print(response['message']['tool_calls'])
More information:
- Python example
- JavaScript example
- API Reference
New models
- Llama 3.1: a new state-of-the-art model from Meta available in 8B, 70B and 405B parameter sizes with support for tool calling.
- Mistral Large 2: Mistral's new 123B flagship model that is significantly more capable in code generation, tool calling, mathematics, and reasoning with 128k context window and support for dozens of languages.
- Firefunction v2: An open weights function calling model based on Llama 3, competitive with GPT-4o function calling capabilities.
- Llama-3-Groq-Tool-Use: A series of models from Groq that represent a significant advancement in open-source AI capabilities for tool use/function calling.
What's Changed
- Fixed duplicate error message when running ollama create
New Contributors
- @lreed-mdsol made their first contribution in https://github.com/ollama/ollama/pull/5757
- @ajhai made their first contribution in https://github.com/ollama/ollama/pull/5799
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.8...v0.3.0

New models
- Mistral Nemo: A state-of-the-art 12B model with 128k context length, built by Mistral AI in collaboration with NVIDIA.
- NuExtract: A 3.8B model fine-tuned on a private high-quality synthetic dataset for information extraction, based on Phi-3.
What's Changed
- Fixed issue where a final assistant message would not be considered for continuing a response
- Improved OpenAI-compatible chat completions endpoint image handling
- ollama create will now validate templates
- Fix error when old versions of ROCm 5 would be detected
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.7...v0.2.8

What's Changed
- Fixed issue where last message when streaming would omit the content response field
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.6...v0.2.7

New models
- Mathstral: MathΣtral is a 7B model designed for math reasoning and scientific discovery by Mistral AI.
What's Changed
- Fixed issue where uppercase roles such as USER would no longer work in the chat endpoints
- Fixed issue where empty system message would be included in the prompt
New Contributors
- @SommerEngineering made their first contribution in https://github.com/ollama/ollama/pull/5721
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.5...v0.2.6

What's changed
- Fixed issue where a model's SYSTEM message not be applied
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.4...v0.2.5

What's Changed
- Fixed issue where context, load_duration and total_duration fields would not be set in the /api/generate endpoint.
- Ollama will no longer error if loading models larger than system memory if disk space is available
New Contributors
- @Jaarson made their first contribution in https://github.com/ollama/ollama/pull/5675
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.3...v0.2.4

What's Changed
- Fix issue where system prompt would not be applied
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.2...v0.2.3

What's Changed
- Fixed errors that occurred when using Ollama with Nvidia V100 GPUs
- glm4 models will no longer fail to load from out of memory errors
- Fixed error that would occur when running deepseek-v2 and deepseek-coder-v2 models
- Fixed a series of out of memory issues when using Nvidia GPUs
- Fixed a series of errors that would occur when using multiple Radeon GPUs
- Fixed rare missing DLL issue on Windows
- Fixed rare crash if ROCm was manually installed on Windows
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.1...v0.2.2

What's Changed
- Fixed issue where setting OLLAMA_NUM_PARALLEL would cause models to be reloaded after each request
Full Changelog: https://github.com/ollama/ollama/compare/v0.2.0...v0.2.1

Concurrency
Ollama 0.2.0 is now available with concurrency support. This unlocks 2 specific features:
Parallel requests
Ollama can now serve multiple requests at the same time, using only a little bit of additional memory for each request. This enables use cases such as:
- Handling multiple chat sessions at the same time
- Hosting a code completion LLM for your internal team
- Processing different parts of a document simultaneously
- Running several agents at the same time.
https://github.com/ollama/ollama/assets/251292/9772a5f1-c072-41db-be6c-dd3c621aa2fd
Multiple models
Ollama now supports loading different models at the same time, dramatically improving:
- Retrieval Augmented Generation (RAG): both the embedding and text completion models can be loaded into memory simultaneously.
- Agents: multiple different agents can now run simultaneously
- Running large and small models side-by-side
Models are automatically loaded and unloaded based on requests and how much GPU memory is available.
To see which models are loaded, run ollama ps:
% ollama ps
NAME ID SIZE PROCESSOR UNTIL
gemma:2b 030ee63283b5 2.8 GB 100% GPU 4 minutes from now
all-minilm:latest 1b226e2802db 530 MB 100% GPU 4 minutes from now
llama3:latest 365c0bd3c000 6.7 GB 100% GPU 4 minutes from now
For more information on concurrency, see the FAQ
New models
- GLM-4: A strong multi-lingual general language model with competitive performance to Llama 3.
- CodeGeeX4: A versatile model for AI software development scenarios, including code completion.
- Gemma 2: Improved output quality and base text generation models now available
What's Changed
- Improved Gemma 2
- Fixed issue where model would generate invalid tokens after hitting context window
- Fixed inference output issues with gemma2:27b
- Re-downloading the model may be required: ollama pull gemma2 or ollama pull gemma2:27b
- Ollama will now show a better error if a model architecture isn't supported
- Improved handling of quotes and spaces in Modelfile FROM lines
- Ollama will now return an error if the system does not have enough memory to run a model on Linux
New Contributors
- @Muku784 made their first contribution in https://github.com/ollama/ollama/pull/5382
- @abitrolly made their first contribution in https://github.com/ollama/ollama/pull/4821
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.48...v0.2.0

What's Changed
- Fixed issue where Gemma 2 would continuously output when reaching context limits
- Fixed out of memory and core dump errors when running Gemma 2
- /show info will now show additional model information in ollama run
- Fixed issue where ollama show would result in an error on certain vision models
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.47...v0.1.48

What's Changed
- Added support for Google Gemma 2 models (9B and 27B)
- Fixed issues with ollama create when importing from Safetensors
A special thank you to the Google Cloud and DeepMind team members for Gemma 2 support.
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.46...v0.1.47

What's Changed
- Increased model loading speed with ollama run, especially if running an already-loaded model
- Improved performance of /api/show including for large models
- Fixes issue where the --quantize flag in ollama create would lead to an error
- Improved model loading times when models would not completely fit in system memory on Linux
- Fixed issue where certain Modelfile parameters would not be parsed correctly
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.45...v0.1.46

New models
- DeepSeek-Coder-V2: A 16B & 236B open-source Mixture-of-Experts code language model that achieves performance comparable to GPT4-Turbo in code-specific tasks.
ollama show
ollama show will now show model details such as context length, parameters, embedding size, license and more:
% ollama show llama3
Model
arch llama
parameters 8.0B
quantization Q4_0
context length 8192
embedding length 4096

Parameters
num_keep 24
stop "<|start_header_id|>"
stop "<|end_header_id|>"
stop "<|eot_id|>"

License
META LLAMA 3 COMMUNITY LICENSE AGREEMENT
Meta Llama 3 Version Release Date: April 18, 2024
What's Changed
- ollama show <model> will now show model information such as context window size
- Model loading on Windows with CUDA GPUs is now faster
- Setting seed in the /v1/chat/completions OpenAI compatibility endpoint no longer changes temperature
- Enhanced GPU discovery and multi-gpu support with concurrency
- The Linux install script will now skip searching for network devices
- Introduced a workaround for AMD Vega RX 56 SDMA support on Linux
- Fix memory prediction for deepseek-v2 and deepseek-coder-v2 models
- api/show endpoint returns extensive model metadata
- GPU configuration variables are now reported in ollama serve
- Update Linux ROCm to v6.1.1
New Contributors
- @jayson-cloude made their first contribution in https://github.com/ollama/ollama/pull/4972
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.44...v0.1.45

What's Changed
- Fixed issue where unicode characters such as emojis would not be loaded correctly when running ollama create
- Fixed certain cases where Nvidia GPUs would not be detected and reported as compute capability 1.0 devices
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.43...v0.1.44

What's Changed
- New import.md guide for converting and importing models to Ollama
- Fixed issue where embedding vectors resulting from /api/embeddings would not be accurate
- JSON mode responses will no longer include invalid escape characters
- Removing a model will no longer show incorrect File not found errors
- Fixed issue where running ollama create would result in an error on Windows with certain file formatting
New Contributors
- @erhant made their first contribution in https://github.com/ollama/ollama/pull/4854
- @nischalj10 made their first contribution in https://github.com/ollama/ollama/pull/4612
- @dcasota made their first contribution in https://github.com/ollama/ollama/pull/4852
- @Napuh made their first contribution in https://github.com/ollama/ollama/pull/4084
- @hughescr made their first contribution in https://github.com/ollama/ollama/pull/3782
- @jimscard made their first contribution in https://github.com/ollama/ollama/pull/3382
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.42...v0.1.43

New models
- Qwen 2: a new series of large language models from Alibaba group
What's Changed
- Fixed issue where qwen2 would output erroneous text such as GGG on Nvidia and AMD GPUs
- ollama pull is now faster if it detects a model is already downloaded
- ollama create will now automatically detect prompt templates for popular model architectures such as Llama, Gemma, Phi and more.
- Ollama can now be accessed from local apps built with Electron and Tauri, as well as in developing apps in local html files
- Update welcome prompt in Windows to llama3
- Fixed issues where /api/ps and /api/tags would show invalid timestamps in responses
New Contributors
- @shoebham made their first contribution in https://github.com/ollama/ollama/pull/4766
- @kartikm7 made their first contribution in https://github.com/ollama/ollama/pull/4719
- @royjhan made their first contribution in https://github.com/ollama/ollama/pull/4822
Full Changelog: https://github.com/ollama/ollama/compare/v0.1.41...v0.1.42

What's Changed
- Fixed issue on Windows 10 and 11 with Intel CPUs with integrated GPUs where Ollama would encounter an error

New models
- Codestral: Codestral is Mistral AI’s first-ever code model designed for code generation tasks.
- IBM Granite Code: now in 3B and 8B parameter sizes.
- Deepseek V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
What's Changed
- Fixed out of memory and incorrect token issues when running Codestral on 16GB Macs
- Fixed issue where full-width characters (e.g. Japanese, Chinese, Russian) were deleted at end of the line when using ollama run
New Examples
- Use open-source models as coding assistant with Continue
New Contributors
- @zhewang1-intc made their first contribution in https://github.com/ollama/ollama/pull/3278

New models
- Cohere Aya 23: A new state-of-the-art, multilingual LLM covering 23 different languages.
- Mistral 7B 0.3: A new version of Mistral 7B with initial support for function calling.
- Phi-3 Medium: a 14B parameters, lightweight, state-of-the-art open model by Microsoft.
- Phi-3 Mini 128K and Phi-3 Medium 128K: versions of the Phi-3 models that support a context window size of 128K
- Granite code: A family of open foundation models by IBM for Code Intelligence
Llama 3 import
It is now possible to import and quantize Llama 3 and its finetunes from Safetensors format to Ollama.
First, clone a Hugging Face repo with a Safetensors model:
git clone https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct
cd Meta-Llama-3-8B-Instruct
Next, create a Modelfile:
FROM .

TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>

{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>

{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>

{{ .Response }}<|eot_id|>"""

PARAMETER stop <|start_header_id|>
PARAMETER stop <|end_header_id|>
PARAMETER stop <|eot_id|>
Then, create and quantize a model:
ollama create --quantize q4_0 -f Modelfile my-llama3
ollama run my-llama3
What's Changed
- Fixed issues where wide characters such as Chinese, Korean, Japanese and Russian languages.
- Added new OLLAMA_NOHISTORY=1 environment variable that can be set to disable history when using ollama run
- New experimental OLLAMA_FLASH_ATTENTION=1 flag for ollama serve that improves token generation speed on Apple Silicon Macs and NVIDIA graphics cards
- Fixed error that would occur on Windows running ollama create -f Modelfile
- ollama create can now create models from I-Quant GGUF files
- Fixed EOF errors when resuming downloads via ollama pull
- Added a Ctrl+W shortcut to ollama run
New Contributors
- @rapmd73 made their first contribution in https://github.com/ollama/ollama/pull/4467
- @sammcj made their first contribution in https://github.com/ollama/ollama/pull/4120
- @likejazz made their first contribution in https://github.com/ollama/ollama/pull/4535

New Models
- Falcon 2: A new 11B parameters causal decoder-only model built by TII and trained over 5T tokens.
- Yi 1.5: A new high-performing version of Yi, now licensed as Apache 2.0. Available in 6B, 9B and 34B sizes.
What's Changed
ollama ps
A new command is now available: ollama ps. This command displays currently loaded models, their memory footprint, and the processors used (GPU or CPU):
% ollama ps
NAME ID SIZE PROCESSOR UNTIL
mixtral:latest 7708c059a8bb 28 GB 47%/53% CPU/GPU Forever
llama3:latest a6990ed6be41 5.5 GB 100% GPU 4 minutes from now
all-minilm:latest 1b226e2802db 585 MB 100% GPU 4 minutes from now
/clear
To clear the chat history for a session when running ollama run, use /clear:
>>> /clear
Cleared session context
- Fixed issue where switching loaded models on Windows would take several seconds
- Running /save will no longer abort the chat session if an incorrect name is provided
- The /api/tags API endpoint will now correctly return an empty list [] instead of null if no models are provided
New Contributors
- @fangtaosong made their first contribution in https://github.com/ollama/ollama/pull/4387
- @machimachida made their first contribution in https://github.com/ollama/ollama/pull/4424

What's Changed
- Fixed issue where models with uppercase characters in the name would not show with ollama list
- Fixed usage string for ollama create
- Fix finish_reason being "" instead of null in the Open-AI compatible chat API.
New Contributors
- @todashuta made their first contribution in https://github.com/ollama/ollama/pull/4362

What's Changed
- Fixed exit status 0xc0000005 error with AMD graphics cards on Windows
- Fixed rare out of memory errors when loading a model to run with CPU

New models
- Llama 3 ChatQA: A model from NVIDIA based on Llama 3 that excels at conversational question answering (QA) and retrieval-augmented generation (RAG).
What's Changed
- Quantization: ollama create can now quantize models when importing them using the --quantize or -q flag:
ollama create -f Modelfile --quantize q4_0 mymodel
[!NOTE] --quantize works when importing float16 or float32 models:
- From a binary GGUF files (e.g. FROM ./model.gguf)
- From a library model (e.g. FROM llama3:8b-instruct-fp16)
- Fixed issue where inference subprocesses wouldn't be cleaned up on shutdown.
- Fixed a series out of memory errors when loading models on multi-GPU systems
- Ctrl+J characters will now properly add newlines in ollama run
- Fixed issues when running ollama show for vision models
- OPTIONS requests to the Ollama API will no longer result in errors
- Fixed issue where partially downloaded files wouldn't be cleaned up
- Added a new done_reason field in responses describing why generation stopped responding
- Ollama will now more accurately estimate how much memory is available on multi-GPU systems especially when running different models one after another
New Contributors
- @fmaclen made their first contribution in https://github.com/ollama/ollama/pull/3884
- @Renset made their first contribution in https://github.com/ollama/ollama/pull/3881
- @glumia made their first contribution in https://github.com/ollama/ollama/pull/3043
- @boessu made their first contribution in https://github.com/ollama/ollama/pull/4236
- @gaardhus made their first contribution in https://github.com/ollama/ollama/pull/2307
- @svilupp made their first contribution in https://github.com/ollama/ollama/pull/2192
- @WolfTheDeveloper made their first contribution in https://github.com/ollama/ollama/pull/4300

New models
- Llava Llama 3: A new high-performing LLaVA model fine-tuned from Llama 3 Instruct.
- Llava Phi 3: A new small LLaVA model fine-tuned from Phi 3.
- StarCoder2 15B Instruct: A new instruct fine-tune of the StarCoder2 model
- CodeGemma 1.1: A new release of the CodeGemma model.
- StableLM2 12B: A new 12B version of the StableLM 2 model from Stability AI
- Moondream 2: Moondream 2's runtime parameters have been improved for better responses
What's Changed
- Fixed issues with LLaVa models where they would respond incorrectly after the first request
- Fixed out of memory errors when running large models such as Llama 3 70B
- Fixed various issues with Nvidia GPU discovery on Linux and Windows
- Fixed a series of Modelfile errors when running ollama create
- Fixed no slots available error that occurred when cancelling a request and then sending follow up requests
- Improved AMD GPU detection on Fedora
- Improved reliability when using the experimental OLLAMA_NUM_PARALLEL and OLLAMA_MAX_LOADED flags
- ollama serve will now shut down quickly, even if a model is loading
New Contributors
- @drnic made their first contribution in https://github.com/ollama/ollama/pull/4116
- @bernardo-bruning made their first contribution in https://github.com/ollama/ollama/pull/4111
- @Drlordbasil made their first contribution in https://github.com/ollama/ollama/pull/4174
- @Saif-Shines made their first contribution in https://github.com/ollama/ollama/pull/4119
- @HydenLiu made their first contribution in https://github.com/ollama/ollama/pull/4194
- @jl-codes made their first contribution in https://github.com/ollama/ollama/pull/3621
- @Nurgo made their first contribution in https://github.com/ollama/ollama/pull/3473
- @adrienbrault made their first contribution in https://github.com/ollama/ollama/pull/3129
- @Darinochka made their first contribution in https://github.com/ollama/ollama/pull/3945

New models:
- Llama 3: a new model by Meta, and the most capable openly available LLM to date
- Phi 3 Mini: a new 3.8B parameters, lightweight, state-of-the-art open model by Microsoft.
- Moondream moondream is a small vision language model designed to run efficiently on edge devices.
- Llama 3 Gradient 1048K: A Llama 3 fine-tune by Gradient to support up to a 1M token context window.
- Dolphin Llama 3: The uncensored Dolphin model, trained by Eric Hartford and based on Llama 3 with a variety of instruction, conversational, and coding skills.
- Qwen 110B: The first Qwen model over 100B parameters in size with outstanding performance in evaluations
- Llama 3 Gradient: A fine-tune of Llama 3 the supports a context window of up 1M tokens.
What's Changed
- Fixed issues where the model would not terminate, causing the API to hang.
- Fixed a series of out of memory errors on Apple Silicon Macs
- Fixed out of memory errors when running Mixtral architecture models
Experimental concurrency features
New concurrency features are coming soon to Ollama. They are available
- OLLAMA_NUM_PARALLEL: Handle multiple requests simultaneously for a single model
- OLLAMA_MAX_LOADED_MODELS: Load multiple models simultaneously
To enable these features, set the environment variables for ollama serve. For more info see this guide:
OLLAMA_NUM_PARALLEL=4 OLLAMA_MAX_LOADED_MODELS=4 ollama serve
New Contributors
- @hmartinez82 made their first contribution in https://github.com/ollama/ollama/pull/3972
- @Cephra made their first contribution in https://github.com/ollama/ollama/pull/4037
- @arpitjain099 made their first contribution in https://github.com/ollama/ollama/pull/4007
- @MarkWard0110 made their first contribution in https://github.com/ollama/ollama/pull/4031
- @alwqx made their first contribution in https://github.com/ollama/ollama/pull/4073
- @sidxt made their first contribution in https://github.com/ollama/ollama/pull/3705
- @ChengenH made their first contribution in https://github.com/ollama/ollama/pull/3789
- @secondtruth made their first contribution in https://github.com/ollama/ollama/pull/3503
- @reid41 made their first contribution in https://github.com/ollama/ollama/pull/3612
- @ericcurtin made their first contribution in https://github.com/ollama/ollama/pull/3626
- @JT2M0L3Y made their first contribution in https://github.com/ollama/ollama/pull/3633
- @datvodinh made their first contribution in https://github.com/ollama/ollama/pull/3655
- @MapleEve made their first contribution in https://github.com/ollama/ollama/pull/3817
- @swuecho made their first contribution in https://github.com/ollama/ollama/pull/3810
- @brycereitano made their first contribution in https://github.com/ollama/ollama/pull/3895
- @bsdnet made their first contribution in https://github.com/ollama/ollama/pull/3889
- @fyxtro made their first contribution in https://github.com/ollama/ollama/pull/3855
- @natalyjazzviolin made their first contribution in https://github.com/ollama/ollama/pull/3962

New models
- WizardLM 2: State of the art large language model from Microsoft AI with improved performance on complex chat, multilingual, reasoning and agent use cases.
- wizardlm2:8x22b: large 8x22B model based on Mixtral 8x22B
- wizardlm2:7b: fast, high-performing model based on Mistral 7B
- Snowflake Arctic Embed: A suite of text embedding models by Snowflake, optimized for performance.
- Command R+: a powerful, scalable large language model purpose-built for RAG use cases
- DBRX: A large 132B open, general-purpose LLM created by Databricks.
- Mixtral 8x22B: the new leading Mixture of Experts (MoE) base model by Mistral AI.
What's Changed
- Ollama will now better utilize available VRAM, leading to less out-of-memory errors, as well as better GPU utilization
- When running larger models that don't fit into VRAM on macOS, Ollama will now split the model between GPU and CPU to maximize performance.
- Fixed several issues where Ollama would hang upon encountering an error
- Fix issue where using quotes in OLLAMA_ORIGINS would cause an error
New Contributors
- @sugarforever made their first contribution in #3400
- @yaroslavyaroslav made their first contribution in #3378
- @Nagi-ovo made their first contribution in #3423
- @ParisNeo made their first contribution in #3436
- @philippgille made their first contribution in #3437
- @cesto93 made their first contribution in #3461
- @ThomasVitale made their first contribution in #3515
- @writinwaters made their first contribution in #3539
- @alexmavr made their first contribution in #3555

New models
- Command R: a Large Language Model optimized for conversational interaction and long context tasks.
- mxbai-embed-large: A new state-of-the-art large embedding model
What's Changed
- Fixed various issues with ollama run on Windows
- History now will work when pressing up and down arrow keys
- Right and left arrow keys will now move the cursor appropriately
- Pasting multi-line strings will now work on Windows
- Fixed issue where mounting or sharing files between Linux and Windows (e.g. via WSL or Docker) would cause errors due to having : in the filename.
- Improved support for AMD MI300 and MI300X Accelerators
- Improved cleanup of temporary files resulting in better space utilization
Important changeFor filesystem compatibility, Ollama has changed model data filenames to use - instead of :. This change will be applied automatically. If downgrading to 0.1.29 or lower from 0.1.30 (on Linux or macOS only) run:

find ~/.ollama/models/blobs -type f -exec bash -c 'mv "$0" "${0//-/:}"' {} \;

New Contributors
- @alitrack made their first contribution in #3111
- @drazdra made their first contribution in #3338
- @rapidarchitect made their first contribution in #3288
- @yusufcanb made their first contribution in #3274
- @jikkuatwork made their first contribution in #3178
- @timothycarambat made their first contribution in #3145
- @fly2tomato made their first contribution in #2946
- @enoch1118 made their first contribution in #2927
- @danny-avila made their first contribution in #2918
- @mmo80 made their first contribution in #2881
- @anaisbetts made their first contribution in #2428
- @marco-souza made their first contribution in #1905
- @guchenhe made their first contribution in #1944
- @herval made their first contribution in #1873
- @Npahlfer made their first contribution in #1623
- @remy415 made their first contribution in #2279
Full Changelog:

AMD Preview
Ollama now supports AMD graphics cards in preview on Windows and Linux. All the features are now accelerated by AMD graphics cards, and support is included by default in Ollama for Linux, Windows and Docker.Supported cards and accelerators
| Family | Supported cards and accelerators |
| -------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- |
| AMD Radeon RX | 7900 XTX 7900 XT 7900 GRE 7800 XT 7700 XT 7600 XT 7600 6950 XT 6900 XTX 6900XT 6800 XT 6800Vega 64 Vega 56 |
| AMD Radeon PRO | W7900 W7800 W7700 W7600 W7500 W6900X W6800X Duo W6800X W6800V620 V420 V340 V320Vega II Duo Vega II VII SSG |
| AMD Instinct | MI300X MI300A MI300MI250X MI250 MI210 MI200MI100 MI60 MI50 |What's Changed
- ollama <command> -h will now show documentation for supported environment variables
- Fixed issue where generating embeddings with nomic-embed-text, all-minilm or other embedding models would hang on Linux
- Experimental support for importing Safetensors models using the FROM <directory with safetensors model> command in the Modelfile
- Fixed issues where Ollama would hang when using JSON mode.
- Fixed issue where ollama run would error when piping output to tee and other tools
- Fixed an issue where memory would not be released when running vision models
- Ollama will no longer show an error message when piping to stdin on Windows
New Contributors
- @tgraupmann made their first contribution in #2582
- @andersrex made their first contribution in #2909
- @leonid20000 made their first contribution in #2440
- @hishope made their first contribution in #2973
- @mrdjohnson made their first contribution in #2759
- @mofanke made their first contribution in #3077
- @racerole made their first contribution in #3073
- @Chris-AS1 made their first contribution in #3094
Full Changelog:

New models
- StarCoder2: the next generation of transparently trained open code LLMs that comes in three sizes: 3B, 7B and 15B parameters.
- DolphinCoder: a chat model based on StarCoder2 15B that excels at writing code.
What's Changed
- Vision models such as llava should now respond better to text prompts
- Improved support for llava 1.6 models
- Fixed issue where switching between models repeatedly would cause Ollama to hang
- Installing Ollama on Windows no longer requires a minimum of 4GB disk space
- Ollama on macOS will now more reliably determine available VRAM
- Fixed issue where running Ollama in podman would not detect Nvidia GPUs
- Ollama will correctly return an empty embedding when calling /api/embeddings with an empty prompt instead of hanging
New Contributors
- @Bin-Huang made their first contribution in #1706
- @elthommy made their first contribution in #2737
- @peanut256 made their first contribution in #2354
- @tylinux made their first contribution in #2827
- @fred-bf made their first contribution in #2780
- @bmwiedemann made their first contribution in #2836
Full Changelog: