Obľúbené: Dáta aktualizované: 19. novembra 2026

Kategórie

Internet1 416
Komunikácia328
Audio a video938
Grafika a dizajn563
Kancelária a produktivita1 179
Vývoj3 284
Umelá inteligencia608
Zabezpečenie408
Systémové nástroje3 187
Hry543
Vzdelávanie a veda242
Sieť a IT293

Strata: staršie verzie na stiahnutie

Runs AI models on your own computer for chat, coding and image work.

Zoznam obsahuje 4 verzie programu Strata – od 9. novembra 2026 po aktuálnu verziu 0.2.16. Každá verzia vedie na inštalátory zo servera výrobcu.

2026: 4 4 26
VerziaDátumVeľkosťSúbory
0.2.16Najnovšia 18. novembra 2026 — 64-bit EXE (NSIS)
0.2.15 16. novembra 2026 — 64-bit EXE (NSIS)
Zoznam zmien
Strata 0.2.15: fixes in the app and the Terminal
The app and the Terminal are 0.2.15. Control is 0.1.61. The engines stay at 0.1.60: the same files as in 0.2.14, which the update feed downloads from the 0.2.14 release. The update feed is sequence 127.
The Intel Arc engine and the DGX Spark engine stay at 0.1.59; the DGX Spark keeps its update feed (sequence 123).
What changes for you
- Nothing to do. Update as usual (see Updating below). Your models, chats and settings stay.
- The engine is not downloaded again: strata update and the app replace the app and the Terminal only.
- There are no new settings.
Fixed
The app
- The app no longer starts llama.cpp's llama-server. In 0.2.14 a model added with "Use a local GGUF…" could load through a llama-server program found on the PC (in STRATA_LLAMA_SERVER, on PATH, or in /opt/homebrew/bin or /usr/local/bin), and the app loaded it again by itself at its next start. Strata now loads models only on its own engine. Such a model shows: "Strata runs models only on its own engine, so this model can't load here. Open Models and choose a model there, or use "Use a local GGUF…" for a file you already have." Checked on the packaged Linux app with a stand-in llama-server named in STRATA_LLAMA_SERVER and first on PATH: 0.2.14 started it after Load (1 start); 0.2.15 started it 0 times, at Load and at its next start.
- Drafting Off in the app no longer downloads the draft layer. In 0.2.14 the app's install asked for the MTP draft layer (about 5 GB) even with Drafting set to Off, on a Mac too. (Checked by the app's tests; not timed on a PC.)
- A Drafting choice made on the app's Models screen is kept after the app restarts. In 0.2.14 the app saved the choice at the start of a Load and wrote the model's settings back without it once the model was ready. After a restart the Models screen showed "Auto (recommended)", and a model set to Off loaded with the DFlash2 drafter. On a Linux PC (Tesla P100, IQ3_XXS) with no MTP draft layer on disk, each choice was made, loaded, and loaded again after an app restart: Off (no drafter, ready in 104 s after the restart), DFlash2 (113 s) and MTP (116 s); the Models screen showed the choice each time.
- The app's note "There is no internet connection, so the draft layer can't be downloaded now" appears only when the model starts without a drafter. In 0.2.14 it could stay on screen after a drafter had been found or downloaded.
- Qwen3.6 35B-A3B on a first run. On a clean Linux PC (2x RTX A6000, 77 MB/s), the first-run install of UD-Q4_K_XL downloaded 23.75 GB and finished in 320 to 340 s, and the bar never stood at 99%. In 0.2.14 this install stopped after 4 seconds.
- The download no longer stops with "this engine (99.0.0) does not run Qwen3.6-35B-A3B yet" while the engine is still being downloaded beside it, and a failed install no longer cancels the engine download.
- Pictures stay on: with no engine on the PC yet, the app now installs the engine first and then the model, so the picture reader (mmproj-F16.gguf, 899,283,584 B) is downloaded with it. On the same PC the install took 350 s, the model loaded with pictures on, and a picture of a red square was answered "It is a red square." (the picture reader started in 0.8 s, using 1,325 MiB of graphics memory).
- The progress bar no longer stops at 99% after the last file. It reaches 100% when every file is in. While the 22.85 GB model file is checked, the bar shows that check ("Preparing files…") instead of a download figure.
- The picture reader's part of the bar counts what has arrived. In 0.2.14 it showed 98 to 99% when 16 to 48% of it had arrived.
- When every file is in but the engine for the PC is not ready, the setup screen now says so: "Your model is downloaded, but the engine for this PC isn't ready yet. Open Settings > About to install it."
The Terminal
- A drafter you choose stays chosen. After a restart without a drafting word, a saved MTP or DFlash2 now holds, as Off already did in 0.2.14.1, and a saved 
0.2.14 15. novembra 2026 — 64-bit EXE (NSIS)
0.2.13 9. novembra 2026 — 64-bit EXE (NSIS)
Zoznam zmien
Strata 0.2.13: new DFlash2 defaults per card, a faster app start, a fix for the Strix Halo Terminal launch, and Qwen4 groundwork
The app and the Terminal are 0.2.13. Control and every engine package are 0.1.59. The update feed is sequence 123.
The DGX Spark (linux-aarch64) gets 0.2.13 too.
Highlights
- DFlash2 drafting, new defaults per card: Strix Halo, UD-Q4_K_XL, 46.6 to 49.1 tokens per second (+5.2%). RTX PRO 5000 72 GB, IQ3_S, 256-261 to 263-265 (+2.8%).
- Strix Halo (Windows): the Terminal launch of UD-Q4_K_XL with pictures off starts. Setup writes a 700 MB reserve, and the engine no longer applies its 2,560 MiB Windows floor on integrated graphics.
- Intel Arc: the same prompt gives the same answer on every run. 50 of 50 runs were identical on an A750, against 3 to 5 different answers in 5 runs before.
- App start: the main file went from 3.12 MB to 0.76 MB, and the start screen from 248.5 ms to 143.4 ms (headless Chrome, median of 10).
- Linux: the expert store loads again when free memory, the file cache or a reserved huge-page pool can hold it (#5327, PR #5531 by pjgmobile).
- Qwen4 groundwork: the signed update feed can carry a model list, and a model of a new shape gets its own engine download beside the Strata engine.
Speed against 0.2.12
Tokens per second, 0.2.12 to 0.2.13, on the DFlash2 defaults that changed. Other machines and models were not measured again.

| Machine, model | Output | Output text |
| ------------------------- | -------------------------- | -------------------------------------------------------------- |
| Strix Halo, UD-Q4_K_XL | 46.6 to 49.1 (+5.2%) | differs from 0.2.12 on 30 of 36 prompts, the same on every run |
| Strix Halo, UD-IQ4_XS-v31 | +2.0% | identical |
| RTX PRO 5000 72 GB, IQ3_S | 256-261 to 263-265 (+2.8%) | identical |

How it was measured: the 0.2.12 settings and the 0.2.13 settings side by side through the Terminal's own flags, 3 interleaved rounds, a fresh engine start per run, greedy answers of 300 tokens. Speeds were not measured again on the final 0.1.59 packages.
What changed for you
- DFlash2 window and confidence per card. Strix Halo: window 5 to 8. Blackwell cc 12.0 with 48 GB or more (the RTX PRO 5000 72 GB class): window 6 to 8, and confidence 0.35 to 0.55 for IQ3_S. The DGX Spark and other cards keep their 0.2.12 values.
- The app's Load fetches the tuned DFlash2 drafter by itself when Auto picks it (as on the RTX PRO 5000 72 GB), as the Terminal does. On an RTX PRO 5000 with IQ3_S (3 rounds of 12 prompts per kind): agent +10.1%, chat +4.8%, all +6.1%, output identical 72 of 72.
- Windows RX 9070 / R9700 (gfx1201) and RX 7900 (gfx1100): the hipBLASLt table is no longer used by default. This restores the 10-19 rule of gfx1151 only, which the engine had overridden since 0.2.10. STRATA_HIPBLASLT_TABLE=on uses the table.
- Settings the launcher used to write are now the engine's own defaults, so a plain engine start matches the app and the Terminal: the parallel stack on Linux CUDA with 2 or more slots, the GDN settings on Linux CUDA, the chunked GDN scan on Linux with all cards gfx12, the gfx1151 hipBLASLt table, the Q8_0 hyper-connection read for Unsloth UD files on CUDA, and fp16 decode windows on the Mac for a resident IQ3_S or Q2_0 pack.
- One fit label in the app and the Terminal: "fits" when the model files without the lookup table fit graphics memory plus RAM, otherwise "may run slower". The Unsloth sets used to be measured against RAM only, which showed "too big" on a V100 32 GB with 31 GB of RAM. DeepSeek's Needs column shows its whole size. NVFP4-QAD is sized at 76.8 GB, the pack that runs.
- Strix Halo (Windows): on UD-Q4_K_XL with pictures off, 22,991 expert-cache slots against 17,556 on 0.2.12, because the engine's reserve floor no longer applies. Configs saved before the fix work after the update.
- Linux: a data folder on an external drive (/run/media/..., /media/..., /mnt) is accepted. Real system folders are still refused. The doubled "Nothing was d

Najobľúbenejšie v kategórii „AI asistenti a LLM“

Zobraziť všetko
  1. Claude CodeAI asistenti a LLM · Anthropic PBC
  2. Codex CLIAI asistenti a LLM · OpenAI, Inc.
  3. ChatGPT by lencxAI asistenti a LLM · lencx
  4. GPT4AllAI asistenti a LLM · Nomic
  5. JanAI asistenti a LLM · Jan
  6. oh-my-piAI asistenti a LLM · Can Bölük
  7. T3 CodeAI asistenti a LLM · T3 Tools
  8. CC SwitchAI asistenti a LLM · farion1231
  9. WSL ManagerAI asistenti a LLM · Bostrot
  10. LobeHubAI asistenti a LLM · LobeHub

Ďalšie programy od výrobcu strata

Zobraziť všetko

Štítky

Zdieľať