LM Studio: versiuni mai vechi, descărcare

- Support for /v1/systemone Jev (TypeSafe AI) compatible endpoint
- Support for /v1/decisions OpenAI compatible endpoint

- [Mac] Added support for the Splash engine on Apple Silicon M3 or newer running macOS 26.4 or newer.

- [LM Studio Engine Protocol] Added advanced llama.cpp argument overrides for GGUF model loading
- Improve DSpark and DFlash assistant drafters compatibility heuristics
- Fixed the Loaded Instances UI showing an incorrect context length
- Fixed /api/v1/chat splitting text and image inputs into separate user messages
- Improved sign-in reliability

- Improve reliability of Qwen 3.8 Flash Next

- Support for DFlash, DSpark, and MTP assistant drafters
- Requires llama.cpp engine version 2.29.1 or newer
- Added support for images returned by tools in OpenAI-compatible /v1/responses and /v1/chat/completions, and Anthropic-compatible /v1/messages
- [Linux] Fixed AppImage startup on systems without FUSE 2
- [Linux] Fixed AppImage startup when Chromium sandboxing is unavailable
- [Linux] Fixed lms wake-up for AppImage installations
- [Linux] Fixed desktop icons for .deb installations
- [Linux] Enabled in-app updates for AppImage and .deb installations
- [Linux] Improved recovery after canceling .deb update authorization
- [Linux] Fixed keyboard shortcuts becoming unavailable after window focus changes on Wayland

- Improved llama.cpp model-load error messages
- Update advanced load settings for mmap, mlock, and direct io for llama.cpp engines 2.28.1 and newer.

- Support for enterprise internal network model endpoint
- Added support for using this device’s models in Bionic over LM Link
- [LM Studio Engine Protocol] Added a developer setting to control the Llama.cpp engine log level

- Add /reasoning to lms chat
- [STABLE] Default LM Studio Engine Protocol to ON.

- Fixed bug where zooming in and out could cause the interface to scale from the top-left and get clipped
- Fixed a bug where reasoning on/off could be ignored by some models with custom reasoning settings
- Fixed automatic chat title generation for some reasoning models
- Added guidance in model settings for llama.cpp options that move when LM Studio Engine Protocol is enabled

- Lm Link no longer requires waitlisting
- Updated default context length to 8k tokens

- Introducing Locally, LM Studio's mobile app. Available on iPhone and iPad.
- Use LM Link in Locally to take your largest LM Studio models on the go
- Security hardening
- [GGUF] Fix multi-GPU selection bugs affecting GPU ON/OFF and Priority Order on some CUDA 12, ROCm, and Vulkan setups

- [CUDA] Added tensor parallelism support for multi-GPU model loading
- [llama.cpp] Added a Physical Batch Size advanced load option
- Fixed REST API requests hanging when HTTP/2 clients sent upgrade headers
- Fixed image attachments appearing in reverse order after sending a chat message
- Fixed a bug where double clicking model trigger would re-open it
- Fixed tool type error bug while using codex --oss
- Fixed invalid role error while using Claude Code. /v1/messages API now supports system messages in the messages array

- Stable release of MTP Speculative Decoding!
- Speeds up generation with models that include built-in multi-token prediction heads
- To try it out, download an MTP-capable model
- Fixed an issue with non-MTP speculative decoding error while MTP was enabled
- Fixed a bug where lms get gemma4 would not show any results
- lms chat now shows which LM Link device each remote model is on

- [MLX] mlx-engine v1.8.1 significantly improves performance and adds parallel predictions for vision-capable models such as Qwen 3.5/3.6 and Gemma 4
- Fixed a bug where newlines were compacted in the chat input on paste
- Bug fixes and security hardening. This update is recommended for all users.

- Support for Qwen 3.6
- Improved style in chat PDF exports
- Fixed bug where MCP servers with OAuth would not work on some Windows environments
- Improved Qwen 3.5 performance with OpenAI-compatible /v1/chat/completions, /v1/responses and Anthropic-compatible /v1/messages

- Support for updated Gemma 4 chat template

- Improve Gemma 4 tool call reliability
- Add OAuth support for MCP servers

- Improve Gemma 4 tool call reliability
- Add support for Anthropic-compatible v1/messages output_config.effort (low, medium, high, max)
- Fixed a bug where deleting a chat folder would sometimes freeze the UI
- Fixed a bug where markdown Link popovers would appear at the top of the window

- Add support for reasoning_effort and reasoning_tokens in OpenAI-compatible v1/chat/completions
- Adds a reasoning field to the /api/v1/models API response, indicating each model's supported reasoning capabilities/REST configuration options learn more
- Fixed a bug where Insert in chat input would sometimes not work after toggling assistant and user mode
- Fixed a bug where surrounding spaces in tool call parameters would be stripped for models that uses XML/XML-like tool call formats
- [CUDA] Fixed issue where some VRAM would not be deallocated under certain conditions
- Fixes a bug where setting reasoning to low when using Nemotron 3 Super via the /api/v1/chat or OpenAI-compatible /v1/responses API would error out

- Fixed a bug when trying to edit an empty chat message

- ✨🎉 Introducing LM Link
- Connect to remote instances of LM Studio, load your models, and use them as if they were local.
- End-to-end encrypted. Launching in partnership with Tailscale.
- Fixed a bug where auto update will sometimes not work due to failing to exit before updater runs.
- This fix only applies, when updating to the next version. That is, you might still encounter issues as you are updating to 0.4.6+1.
- Fixed Qwen3.5 RAG jinja rendering bug: "No user query found in messages"
- Updated go version to 0.25.7 for LM Link.
- [DGX Spark] Enable Direct I/O to improve model load latency
- Requires llama.cpp engine 2.5.1 or greater.

- Fixed a bug where LM Link connector was not included in in-app updater

- Fixed a bug where thinking tags were not being emitted correctly through v1/chat/completions OpenAI-compatible REST API
- Fixed a bug where RAG will sometimes get stuck in "Deciding how to handle the document(s)..."
- Fixed a bug where previously detected GGUF models that were not being properly identified are now correctly recognized

- Fixed bug causing TypeError: Cannot read properties of undefined (reading 'backendInfo')

- Introducing Parallel Requests with MLX! 🎉
- Added support for continuous batching in mlx-engine 1.0.0
- At the moment this capability is text-only, with VLM support in the works
- Fixed a scroll bug on plugin lists
- Fixed a responsive layout bug when role and insert buttons were displayed in chat input
- Introduced new styling for message file attachments

- New: Anthropic API compatibility endpoint: /v1/messages
- Use Claude Code with LM Studio 👾
- Added a new "Deep Dark" theme option for the UI
- Added --parallel <N> flag to lms load command to allow loading a model with specified number of max parallel predictions
- [Mac] Fixed "kernel_mul_mv_bf16_f32_4 was not found in the library" bug on model load
- Fixed a bug where lms chat cannot use non-ASCII characters
- Fixed a bug where UI will sometimes crash after a loaded model is removed
- Fixed a bug where LM Studio sometimes creates multiple backups for model index cache
- Fixed responsive handling of chat terminal icons
- Fixed a bug causing UI to crash with "Cannot read properties of null"
- Fixed a bug to persist expanded config section choices across sessions
- Fixed a bug in splitting and collapsing panes, and replacing chat tabs
- Fixed a bug where Codex would encounter errors when unsupported tools were passed to LM Studio
- Fixed a bug where the model loader would sometimes not populate with correct override model default values
- Fixed a bug where input_tokens and cached_tokens were sometimes reported incorrectly in /v1/responses API
- Fixed edge cases in which Developer Mode wouldn't turn on automatically when updating 0.3 -> 0.4

0.3.39 - Release Notes
Build 1
- Support for image_url input in OpenAI-compatible v1/chat/completions REST endpoint
- Support for top_logprobs in OpenAI-compatible v1/responses REST endpoint
- Output cached_tokens statistics in OpenAI-compatible v1/responses REST endpoint

0.3.37 - Release Notes
Build 1
- Support for the LFM2 tool call format
- Fix "Cannot read properties of null (reading 'architecture')" when using a generator

0.3.36 - Release Notes
Build 1
- FunctionGemma Support

0.3.35 - Release Notes
Build 1
- [MLX] Support for Devstral-2 and GLM-4.6V
- Fixed a bug where the default system prompt was still sent to the model even after the system prompt field was cleared.
- Fixed a bug where exported chats did not include the correct system prompt.
- Fixed a bug where the token count was incorrect when a default system prompt existed but the system prompt field was cleared.
- Fixed a bug where sometimes the tool call results were not being added to the context correctly

0.3.34 - Release Notes
Build 1
- Support for EssentialAI's rnj-1 model
- Fix jinja prompt formatting bug for some models where EOS tokens were not being included properly

0.3.33 - Release Notes
Build 1
- Support for MistralAI's Ministral models (3B, 8B, 13B)
- Support for Olmo-3 tool calling

0.3.32 - Release Notes
Build 2
- Support for GLM 4.5 tool calling
- [MLX] Fixed prompt template bug that caused GLM-4.1V to not recognize images
- Support for olmOCR-2
- Fix a bug where sometimes Download button would continue showing for an already downloaded model
Build 1
- Support for passing base64 images into OpenAI-compatible /v1/responses endpoint
- See https://platform.openai.com/docs/guides/images-vision?api-mode=responses&format=base64-encoded#giving-a-model-images-as-input for details
- Flash Attention is now enabled by default for Vulkan and Metal llama.cpp engines
- Fix OpenAI-compatible /v1/responses endpoint "previous_response_not_found" bug due to internal file read error
- Fix a bug where update toast would sometimes retrigger and close
- Fix the "No model selected for this chat and no lastUsedModel recorded. Please select a model" error
- Fixed cases where downloading additional variants of the same model sometime wouldn't get nested correctly

0.3.31 - Release Notes
New in LM Studio 0.3.31:
- Improve image understanding output quality, especially for OCR tasks
- Default image attachment size is now 2048px on the longest side for better quality with vision models
- Can be changed in Settings > Chat > Image Inputs
- Support for MiniMax M2 tool calling
- Flash Attention is now enabled by default for CUDA engines
- New CLI commands to manage engines:
- lms runtime get
- lms runtime update
- Improved macOS 26 compatibility and support
Build 7
- Fixed a bug where sometimes the Load Model button was greyed out for large MLX models
- Set Flash Attention on by default for CUDA engines
Build 6
- Added tool use support for MiniMax M2
- Added better controls for image input size for vision models
Build 5
- Update dependencies to support macOS 26
- Server UX improvements
- Log v1/chat/completions prompt processing progress to Developer Logs
- Model-specific v1/chat/completions Developer Logs
- Log error when JIT model load failed due to guardrails settings
Build 4
- Fixed "vision.imageResizeSettings key does not exist" error
- Added settings for configuring maximum image input dimensions for vision models.
- Accessible via Settings > Chat > Image Inputs > "Image resize bounds"
- These settings control the maximum dimensions to which input images are resized before being sent to vision-capable models.
Build 3
- Added lms runtime get and lms runtime update CLI commands to manage runtime extensions from the terminal.
- Run lms runtime -h for more info.
- Fix issue where sometimes reasoning models' output would start with a plaintext <think> token
Build 2
- Increase image resize limits to 1024x1024 for improved image processing performance (superseded: now 2048px)
- In a future update this will be user configurable (fixed in build 4)
Build 1
- [MLX] Fix ValueError: Image features and image tokens do not match for Qwen3 VL Thinking models
- Fix occasional UI crash when searching models

0.3.30 - Release Notes
Build 1
- Fixed streaming mode bug that was impacting tool calling functionality for Qwen 3 Coder in /v1/chat/completions API.
- [Vulkan] Fix bug where models were not being loaded onto iGPUs (requires runtime update)
- Compatibility support for the developer role in /v1/responses API endpoint. For now, developer messages will be processed as system messages internally.

0.3.28 - Release Notes
Build 2
- Fix displaying model parameter counts in My Models and model picker
- Improved Qwen format tool call reliability with llama.cpp engines
- Fix UI theme resetting to 'Auto' after update to 0.3.28 (fixes issue in build 1)
Build 1
- New in My Models: ability to quickly select which model variant will be used by default
- Significantly improved RAM/VRAM usage estimation for models that have mixed quantizations
- Significantly improved RAM/VRAM usage estimation for models that have inaccurate parameter labels in their metadata
- Significantly improved RAM/VRAM usage estimation for non-transformer models
- Fixed a bug where models loaded via the CLI would not respect global GPU settings
- Fix some flickering UI interactions on My Models page, and collapse indexer warnings under a new status button
- Fix a bug where citations count badge with over 10 would clip text
- Fix inconsistency in quantization badge text alignment

0.3.27 - Release Notes
Build 4
- Improved VRAM usage estimation, especially when flash attention is enabled
Build 3
- Setting to control whether to open the downloads panel after starting a model download (default: false)
- Update CLI (lms) output colors to have better contrast in light mode
- Fix a bug where copy buttons would sometimes not appear on conversation code blocks
Build 2
- New: Find in Chat (Cmd/Ctrl+F) and Search All Chats (Cmd/Ctrl+Shift+F).
- New: Sort chats sidebar by date updated, date created, or token count
- Model resources estimation will now work for vision models
- Added model resources estimation to the CLI. You can now run lms load --estimate-only <model-name> to preview a model's estimated memory requirements before loading
- While using lms chat, you can now use Ctrl^C to interrupt ongoing predictions
Build 1
- Additional model quantization files downloaded after the main model will now be properly nested under it
- Improved memory usage estimation used for model loading guardrails.
- Now memory estimation will take into account selected context lengths.
- lms ps --json now reports model generation status and the number of queued prediction requests

0.3.26 - Release Notes
Build 6
- The LM Studio CLI (lms) now supports streaming server logs, as well as model output.
- Use lms log stream --source server for server logs
- Use lms log stream --source model --filter input,output for both model input and output logs
- Append --json to get JSON-formatted logs
- Learn more: https://lmstudio.ai/docs/cli/log-stream
- Fixed a bug where clicking Eject in the developer page would sometime open the configuration panel as well
Build 5
- [Windows] fixed a bug where Mission Control buttons were hard to click / behaved like a dragging surface
Build 4
- Fixed a bug where it was sometimes not possible to drag the app action bar when there was an image under it
- New context menu 'Enclose in Folder' options when selecting multiple chats
Build 3
- [Linux] Fixed bug where rag-v1 did not work due to a missing embedding model
- [UI] Switch to native context menus
Build 2
- Restored lms ls --detailed for backwards compatibility. Use lms ls or lms ls --json instead
- Fixed bug where child processes were not cleaned up after LM Studio received SIGKILL
Build 1
- [MLX] Add badge for MXFP4 quantization type

0.3.25 - Release Notes
New in LM Studio 0.3.25:
- Select multiple chats in the sidebar for quick bulk actions (move, delete)
- New option to move deleted chats to trash bin (or: instant delete)
- NVIDIA Nemotron-Nano-v2 support
Build 2
- NVIDIA Nemotron-Nano-v2 support (incl. tool calling)
Build 1
- Select multiple chats and folders by holding Ctrl on PC or Cmd on Mac. Hold shift + click to select a range.
- Option to move items to the operating system's trash bin instead of instantly deleting them
- Improve app settings style and legibility
- When selecting a chat that uses a different model, auto switch to that model if it's loaded
- Bug fix: new chats and folders now correctly created relative to the selected folder
- Chat: moved token counters to the app footer
- Bug fix: ensure copy button appears on plaintext code blocks

0.3.24 - Release Notes
New in LM Studio 0.3.24:
- 🎉 Support for ByteDance/Seed-OSS (36B)
- Improved code blocks (sticky copy button) and markdown tables style
- Improved lms (LM Studio CLI) command output style
- Bug fixes
Build 6
- Fix a bug where markdown blocks could overflow beyond the chat screen
- Fixed an issue with Seed-OSS where LM Studio did not parse tool calls for certain prompt templates
Build 5
- Support for ByteDance/Seed-OSS, including tool calling support and prompt template fixes
- lms ps now provides more information about model status, context and TTL.
- Fix a bug where pressing Tab did not correctly indent within the system prompt editor (Cmd/Ctrl + E)
Build 4
- Fixed a UI crash that would happen if the model being loaded was larger than the user's guardrails
- Make lms (LM Studio CLI) output style more consistent
Build 3
- Fixed a bug where multiple config layers in model.yaml were in the wrong order
- Added ability to edit chat name from chat tab dropdown, or double clicking chat tab
Build 2
- Fix sticky copy button bottom padding
- Fix setting numCpuExpertLayersRatio via the SDK
Build 1
- Fix bug where context menu click on model loader would cause a flicker
- Improved styling for tables in markdown
- Improve code blocks: polished style and sticky copy code button
- Fix bug where Harmony updates didn't respect the runtime extension pack "auto-update" setting
- Adds Gaeilge translation, thanks to [@aindriu80] (https://github.com/aindriu80)
- Updates to translations of Catalan, Chinese (Simplified), Chinese (Traditional), Japanese and Portuguese (Brazil).
- Added app setting for default context length for all models. See under App Settings > General
- Fixed a bug where setting context length to 0 would cause an error
- Fixed a bug where setting values between (1.0, 1.1) would not work
- Copying an assistant message now only copies the content, without formatting or thinking tokens
- Fixed a bug where leftover speculative decoding setting would disable Markdown in chat

0.3.23 - Release Notes
Build 3
- [llama.cpp][MoE] Add ability to offload expert weights to CPU/GPU RAM via "Force Model Expert Weights onto CPU" in advanced load settings
- Tool names are normalized before being provided to the model (replace whitespaces, special chars)
Build 2
- Fix "Complete Download" button sometimes not working when downloading a staff-picked model
- Fix "Fix" button not working for extension packs (like Harmony)
- Fix "Cannot read properties of undefined (reading 'properties')" for certain tools-containing requests to /v1/chat/completions
- Fix Error: EPERM: operation not permitted, unlink when auto-updating harmony
Build 1
- Bug fixes resulting in significant improvements for in chat tool calling reliability
- Fixed a bug where some old conversations won't load in the app
- Fixed a bug where tool call will fail sometimes when used via OpenAI compatible API in non-streaming mode
- Fix models not outputting thinking tags in v1/chat/completions
- For gpt-oss:
- message.content will not include reasoning content or special tags
- This matches the behavior of o3-mini.
- Reasoning content will be in choices.message.reasoning (stream=false) and choices.delta.reasoning (stream=true)
- Fix "TypeError: Invalid Version" causing app functionality issues on machines with AMD+NVIDIA GPUs
- Fix bug where MCP plugin chip name was not rendering in User Mode
- Fix bug where search results were refreshing on click

0.3.22 - Release Notes
New in LM Studio 0.3.22:
- 🎉 Support for OpenAI's gpt-oss models!
- We partnered with OpenAI to bring their open-source models to LM Studio on launch day.
- There are 2 models: gpt-oss-20b and gpt-oss-120b.
- Learn more: https://lmstudio.ai/blog/gpt-oss
Build 1
- [gpt-oss] Bundle OpenAI/Harmony parsing framework on all platforms
- [gpt-oss] More forgiving parsing of tool calls in chat
- [gpt-oss] [OpenAI-compat API] Remove structural tokens (<|channel|>, <|end|>, etc...) from v1/chat/completions output
- Should address issues seen with tools such as OpenCode

0.3.21 - Release Notes
Build 4
- Simplify Mission Control for User Mode
Build 3
- Parse Qwen3 coder tool calls that start with <function=... and not <tool_call>
Build 2
- Fix Qwen3-235B-A22B-Thinking-2507 thinking block handling
- Fix Qwen3 coder 'tool_calls' emission in OpenAI-like /v1/chat/completions server
- Introduce a Settings > Chat option to hide token counts from chat cells in the sidebar
- Fix a bug where keyboard shortcut to chats page was lost after switching app modes
- Avoid auto dismissing toasts if currently hovered over
- Fix a bug where toggling message style would not reinitialize expandable messages
- For new users, default to not showing chat token count in chat sidebar cells
- Fixes 'Cannot apply filter "string" to type: NullValue' with Qwen3 Coder MOE models where null values were not being handled correctly in jinja parsing
Build 1
- Add confirmation toast after adding an MCP server
- Show streaming preview for generating tool call arguments
- By default, filter Staff Picks to models that can fit on your device

0.3.20 - Release Notes
Build 3
- Qwen3-Coder-480B-A35B tools support
Build 2
- Fixed a bug where the app would fail to open
Build 1
- Fix model loader flicker after selecting a model

Build 2
- Fixes to in-app update when service mode is enabled
- Fixes to JIT loading across client application lifetime
Build 1
- Run LM Studio as a service (headless)
- lms load, lms server start no longer requires launching the GUI
- ability to run on machine startup
- Server start / stop button will remember last setting
- This is useful when LM Studio is running as a service
- Improvement to Model Search
- Hugging Face search now happens automatically without Cmd / Ctrl + Enter
- Just-In-Time model loading for OpenAI endpoints
- Button to toggle Mission Control full screen / modal modes
- Update llama.cpp-based JSON response generation; now supports more complex JSON schemas
- Tray menu options to minimize app to tray, copy server base URL
- Checkbox to add lms to PATH during onboarding on Linux
- [Mac][MLX Vision] Bump mlx-vlm version to 0.0.15, support Qwen2VL
- [Mac][MLX Engine] Updated Transformers to 4.45.0
- Fixes some issues with sideloading quantized MLX models (https://github.com/lmstudio-ai/mlx-engine/issues/10)
- [UI] Move Chat Appearance control to top bar
- [UI] Tweaks to size of per-message action buttons
- Localization:
- Improved German translation thanks to Goekdeniz-Guelmez
- Indonesian translation thanks to dwirx
Bug fixes
- [Bug fix] fix RAG reinjecting document into context on follow up prompts
- Fixed RAG not working (https://github.com/lmstudio-ai/mlx-engine/issues/4)
- Fix outline flicker around Mission Control

What's new in 0.3.4
New
- [Mac] Support for MLX! Read the announcement here
- Mission Control: Cmd+Shift+M to search for models, Cmd+Shift+R to manage LM Runtimes
- Set Structured Output (JSON Schema) from the UI
- [UI] Move Chat Appearance control to top bar
- [UI] Tweaks to size of per-message action buttons
Bug Fixes
- Fix for Black Screen after prolonged use (reference: lmstudio-bug-tracker#98)
- Fix for no port other than 1234 working for the local server (reference: lms#80)
- Fix for embedding API not working from Obsidian (reference: tracker#142)
- Fix for RAG sometimes failing during document processing

What's new in 0.3.2
- New: Ability to pin models to the top is back!
- Right-click on a model in My Models and select "Pin to top" to pin it to the top of the list.
- Chat migration dialog now appears in the chat side bar.
- You can migrate your chats from pre-0.3.0 versions from there.
- As of v0.3.1, system prompts now migrated as well.
- Your old chats are NOT deleted.
- Don't show a badge with a number on the downloads button if there are no downloads
- Added a button to collapse the FAQ sidebar in the Discover tab
- Reduced default context size from 8K to 4K tokens to alleviate Out of Memory issues
- You can still configure any context size you want in the model load settings
- Added a warning next to the Flash Attention in model load settings
- Flash Attention is experimental and may not be suitable for all models
- Updated the bundled llama.cpp engine to 3246fe84d78c8ccccd4291132809236ef477e9ea (Aug 27)

Bug Fixes
- Bug fix: My Models model size aggregation was incorrect when you had multi-part model files
- Bug fix: (Linux) RAG would fail due to missing bundled embedding model (fixed)
- Bug fix: Flash Attention - KV cache quantization is back to FP16 by default
- In 0.3.0, the K and V were both set to Q8, which introduced large latencies in some cases
- You might notice an increase in memory consumption when FA is ON compared with 0.3.1, but on par with 0.2.31
- Bug fix: On some setups app would hang at start up (fix + mitigation)
- Bug fix: Fixed an issue where the downloads panel was dragged over the app top bar
- Bug fix: Fixed typos in built-in code snippets (server tab)