unslothai/unsloth v0.1.812-beta on GitHub

We're releasing support for Qwen-Image-2.1, custom Agent Skills, and easier chat/project management. It also brings 2x faster reasoning blocks (60 FPS vs 30 FPS), more reliable training, and improved Linux installs and updates. Qwen Image 2.1 Guide

Highlights

Qwen-Image-2.1 support for image generation and more
Add custom skills to guide models through specific tasks.
2x faster long reasoning blocks (60 FPS vs 30 FPS before)
In-app Debian updates and an Ubuntu 24.04+ installer for ARM64.
Chats are now fully draggable in sidebar (web UI + Linux desktop app for now, full support coming very soon). Thinking UI/UX reworked.
Projects are easier to edit and organize.
In-app Debian updates and an Ubuntu 24.04+ installer for ARM64.

Agent Skills

Add custom skills to guide models through specific tasks.
Reuse skills from your existing Claude Code and .agents folders.
Manage your skills and use @ to select one in chat.

Chats + projects

Drag chats to reorder them, pin them or move them into a project.
Better model loading behavior, chat properly remembers settings
Thinking UI/UX reworked for cleaner modern experience.
Edit project names, instructions and folders directly from the Projects page.
Long code replies stream more smoothly while keeping syntax highlighting.

Models + storage

Vision models can now see images returned by MCP tools.
Other accounts can connect to local model servers when the owner allows it.
Chat can reuse downloaded models instead of fetching a second copy.
CLI chats preserve loaded models’ precision and use recommended settings.
See which caches take up disk space and clear them from Settings.

Training + exports

More reliable training resumes and dataset preparation.
Model uploads to the Hub leave out files from earlier exports.
Hugging Face datasets now work in recipes, past training runs can be resumed, resumed runs show the correct remaining time, dataset columns are mapped correctly, and selected splits/subsets are respected.

Desktop + hardware

Linux/Desktop distribution got substantially better - there are now in-app updates for Debian installs and a native ARM64 Linux .deb.
New AMD ROCm Docker image with Unsloth Studio and JupyterLab.
Larger text now has more room, and contrast settings make buttons and menus easier to see.
Filter logs, wrap long lines and jump to the latest entries without scrolling through everything.

What's Changed

Bump install.sh / install.ps1 pin to unsloth>=2026.9.7 by @danielhanchen in #11279
Carry over five temp-scanner fixes that landed after #11217 merged by @danielhanchen in #11280
fix the Hub README's tag and GPU support claims by @NilayYadav in #11131
Give the page-fault stall test a window the scheduler cannot close by @danielhanchen in #11284
Stop three tests stranding a waiting request in a module-level dict by @danielhanchen in #11287
Assert the companion scan's thread, not how fast the machine is by @danielhanchen in #11288
studiobench: count the attachments the composer actually renders by @danielhanchen in #11302
Route today's new PowerShell test calls through the shared runner by @danielhanchen in #11285
Studio: run one repair at a time, however many times Retry is pressed by @LeoBorcherding in #11249
Route the last direct pwsh spawns through the shared runner by @danielhanchen in #11303
Studio: keep a LoRA push private when the Hub repo already exists by @NilayYadav in #11229
Studio: preserve Claude Code prompt cache with custom output styles by @Imagineer99 in #11121
Disengage FBCache on the word of whoever owns it by @danielhanchen in #11304
Studio: preserve partial state for canceled GGUF companions by @Etherll in #11018
Studio: restore an exported chat backup as separate chats by @NilayYadav in #11231
Studio: fetch a remote image URL instead of forwarding it to llama-server by @oobabooga in #11160
Studio: tell the model which tool calls were skipped by the per turn limit by @NilayYadav in #11154
fix(desktop): explain updates blocked by an existing Studio server by @Imagineer99 in #11273
Desktop: support SQF script attachments by @Imagineer99 in #11274
Studio: preserve local model inventory after cache filesystem errors by @Imagineer99 in #11272
Studio: train a gated upstream model through its public Unsloth copy by @oobabooga in #11214
Stop one test's app settings answering another test's read by @danielhanchen in #11309
Make the WSL compiler-intermediate control say why it failed by @danielhanchen in #11312
Studio: let models see images returned by MCP tools by @NilayYadav in #10088
Studio: trim the composer "+" menu top level by @shimmyshimmer in #11314
Studio: optically centre the model selector label by @shimmyshimmer in #11316
Put back the Docker quickstart paragraph the README lost by @danielhanchen in #11317
Re-approve three unsloth-zoo exec sites, pinned, and say which reviews are which by @danielhanchen in #11318
Serve the decoded pixels on a multi-image turn, not the base64 they arrived as by @danielhanchen in #11320
Clear the idle-reload stash between tests by @danielhanchen in #11322
Desktop clean machine: pick the x64 installer, and stop blaming the .deb for the mirror by @danielhanchen in #11323
Acknowledge the Colab oracle drift, and stop the diff hiding entries by @danielhanchen in #11324
Give the torchcodec placeholder a real spec by @danielhanchen in #11325
Strip distro .devN markers from the Colab pins by @danielhanchen in #11326
Give the CUDA spoof the raw-stream handle unsloth reads at import by @danielhanchen in #11328
Studio: report a backend that is provably gone without sleeping out the retry ladder by @danielhanchen in #11213
Name an AMD device node this account cannot open, instead of reading it as no GPU by @danielhanchen in #10473
Studio: keep a connected pin that could not be written by @shimmyshimmer in #11269
Studio setup: reuse the uv a previous run installed instead of downloading it again by @danielhanchen in #10659
Point the interrupt legs at the phase label the installer prints by @danielhanchen in #11342
Studio: tidy run settings controls and settings page titles by @shimmyshimmer in #11331
Cancel superseded pull request runs, and guard that they stay cancelled by @danielhanchen in #11345
Format the superseded-runs guard the way the hook does by @danielhanchen in #11348
Studio: tidy the run settings panel by @shimmyshimmer in #11346
Studio: let recipes with a Hugging Face dataset run by @NilayYadav in #11300
Studio: bring back the Resume button for past training runs by @NilayYadav in #11301
Studio: map context, question and answer columns to the right roles by @NilayYadav in #11292
Studio: show the right time left after resuming training by @NilayYadav in #11297
Say which locale catalog failed instead of dying in an eval by @danielhanchen in #11344
Studio: push only the exported files to the Hugging Face Hub by @NilayYadav in #11299
Retry a browser driver once before failing the leg by @danielhanchen in #11355
Take the Linux subdirectory-watch race out of the compiler watcher test by @danielhanchen in #11362
Docker ROCm image: reach the GPU through WSL2's DXG bridge, not only /dev/kfd by @LeoBorcherding in #11212
Studio: turn on Code when Full access is picked by @oobabooga in #11283
Studio: chat with a cached upstream model instead of downloading its Unsloth copy by @oobabooga in #11307
studio: add in-app updates for debian installs by @mahiatlinux in #11338
Let the uv-safe requirements test pass where 8.3 names exist by @danielhanchen in #11365
studio: protect authenticated release redirects by @mahiatlinux in #11339
Studio: give a project one home in the sidebar, and fix the row menus around it by @shimmyshimmer in #11370
Studio: reopen dismissed Deep Research plan dialogs by @AzizMuminov in #10802
desktop: publish a native arm64 linux deb by @mahiatlinux in #11356
Studio: export and save only the reply you are looking at by @NilayYadav in #11298
fix(studio): track Marked's def rule in the link definition probe by @HughScott2002 in #9645
Studio: skip duplicate weight files when downloading a model from the Hub by @NilayYadav in #11227
Docker ROCm image: ship Unsloth Studio, as the CUDA image already does by @LeoBorcherding in #11218
Keep the regenerable caches inside the Studio root by @danielhanchen in #10597
Studio: show Vulkan iGPU memory as shared availability by @oobabooga in #11305
Studio: put the sidebar nav validator signature on the line the formatter wants by @danielhanchen in #11374
Studio: drop rows the chat template can't format and tell the user by @NilayYadav in #11293
Let the pinned formatter have the line it wants in settings.py by @danielhanchen in #11371
Give the desktop release tests an explicit utf-8 encoding by @danielhanchen in #11372
Correct why a superseded probe matrix is expensive: slots, not minutes by @danielhanchen in #11347
Say when a failing pull_request job is testing a stale merge ref by @danielhanchen in #11386
Read the routing invariance counters as a steady state, not one request by @danielhanchen in #11389
Hold the diffusion load worker on an event instead of a sleep by @danielhanchen in #11392
Studio: drag sidebar rows into place, and make a project row do more by @shimmyshimmer in #11373
Studio: use only the chosen split and subset for Hugging Face recipe data by @NilayYadav in #11294
Count a unified-memory APU's window as shared host memory, not dedicated VRAM by @danielhanchen in #11366
Studio: use copy-01 for every copy button by @shimmyshimmer in #11395
Studio: treat bash coproc as a command boundary in the terminal safety scanners by @danielhanchen in #11369
Drain worker threads without racing their start by @danielhanchen in #11400
Studio: read the local model cache locally, and keep host paths off the API by @danielhanchen in #11026
fix(studio): run built-in text search as a tiered engine allowlist by @alkinun in #11242
Give a mixed NVIDIA plus AMD host a way to ask for the ROCm torch stack by @danielhanchen in #10474
Click the reasoning trigger by slot, not by its wording by @danielhanchen in #11406
Studio: use one expand icon and one download icon everywhere by @shimmyshimmer in #11411
Studio: one white and one grey for dark mode thread text by @shimmyshimmer in #11410
Studio: round the right corners of scrollable tool result and thinking panes by @shimmyshimmer in #11409
Studio: rename the edit composer's Update button to Send, and resend an unchanged prompt by @shimmyshimmer in #11407
Studio: answer a recipe check the backend cannot read the seed for by @danielhanchen in #11363
Repair the two frontend CI breaks main inherited from #11373 by @danielhanchen in #11416
Studio: bind the Data Recipe export link to the account that minted it by @danielhanchen in #11401
Windows installer: stop an unresolvable 8.3 alias aborting the rollback by @umran666 in #11319
Studio: make the rule closing a thinking trace readable by @shimmyshimmer in #11424
Studio: one rounded outline on the guided tour card by @shimmyshimmer in #11421
Studio: brighten tool rows on hover, like the Thinking trigger by @shimmyshimmer in #11422
Studio: use the scroll icon for Agent Skills everywhere by @shimmyshimmer in #11419
Studio: say what auto-compaction actually does by @shimmyshimmer in #11415
Studio: put a working chat row's spinner back on the trailing column, and tighten the pin gap by @shimmyshimmer in #11408
Studio: icon-only Copy and Download in the transcript, with tooltips by @shimmyshimmer in #11425
Use one speech bubble for a chat everywhere by @shimmyshimmer in #11428
Studio: call the product Unsloth in the app's own copy by @shimmyshimmer in #11427
Scale the settings hint icon with the UI font size by @shimmyshimmer in #11429
Studio: fall back to the Vulkan sd.cpp build when the ROCm one cannot run by @danielhanchen in #11068
Read only the selector shapes the sheet test can prove, and refuse the rest by @danielhanchen in #11089
Make the GRPO canary run through Unsloth's own batch sampler by @danielhanchen in #11126
Studio: call it Skills, and lowercase files in Chat with files by @shimmyshimmer in #11426
Stop the barrel guard counting package imports as a way back in by @shimmyshimmer in #11431
Reject a new studio union offender even when the recorded count is unchanged by @danielhanchen in #9856
Derive the chat UI wall-clock watchdog from the turn timeout by @danielhanchen in #9025
Studio: scope the GGUF converter pin to the conversion by @danielhanchen in #11417
Studio auth and bootstrap gates: throttle desktop-login, narrow the tunnel CORS origin, identify Colab's proxy by @danielhanchen in #11398
fix(studio): widen the link reference probe to match the definition probe by @danielhanchen in #11375
remove gguf shard controls introduced by #6107 by @mahiatlinux in #11315
Fail Source lint on a duplicate TypeScript import binding by @danielhanchen in #8548
A failed presence check in the indicator driver should say what the page looked like by @danielhanchen in #10435
Studio: the drag-and-drop switches go, the gesture has one behaviour by @shimmyshimmer in #11441
Studio: drop the "When context fills" setting and always use the server's policy by @shimmyshimmer in #11444
Studio: read the saved credential and its presence in one query by @danielhanchen in #11420
Studio: require a published digest on upstream llama.cpp prebuilts too by @danielhanchen in #11402
fix(sft): honor max_length below the model context limit by @taking-lying-flat in #11266
Studio: three-state Thinking and Tool calls visibility in Display by @shimmyshimmer in #11433
Refuse a compiler cache path shlex cannot parse, not only one with a space by @danielhanchen in #11414
[perf] Qwen3.5-9B LoRA SFT with unsloth-cli.py on a B200: 0.84 s to 0.77 s per step at 1 GPU, the model-independent half of #10744 by @TarzanZhao in #11238
Offer FlexAttention when head_dim puts every flash kernel out of reach by @danielhanchen in #11102
Studio: edit a project from the Projects page, and sort it from the Updated column by @shimmyshimmer in #11443
Repair four checks that are red on main itself by @danielhanchen in #11456
studio: preserve standalone gguf settings in compare mode by @mahiatlinux in #11311
Stop SFTTrainer downcasting configs that subclass SFTConfig by @danielhanchen in #11440
Point the two compiler-cache fallback tests at a temporary root Windows accepts by @danielhanchen in #11455
Docker ROCm Studio image: ship JupyterLab with the notebooks, as the CUDA image does by @LeoBorcherding in #11286
Studio: prefer the discrete AMD GPU when an iGPU enumerates first on Linux by @danielhanchen in #11437
Studio: let the owner allow managed accounts private provider base URLs by @danielhanchen in #11436
Studio: keep an AMD dGPU and a Vulkan choice through install and update by @danielhanchen in #11442
Apply the logit scale on the fused cross entropy path by @vineethsaivs in #11277
Studio: stop suggesting bnb-4bit models MLX cannot load on Mac by @NilayYadav in #9722
Read the sidebar and reasoning contracts as claims, not as literals by @danielhanchen in #11404
Studio: isolate the sandbox traversal test from the real home directory by @danielhanchen in #11471
Stop four POSIX-only test modules from failing to collect on Windows by @danielhanchen in #11457
Studio: keep the account confinement tests off the interpreter walk by @danielhanchen in #11488
Studio: pin the hardware verdict in the two export capability tests that depend on it by @danielhanchen in #11490
studio: copy full prompts from the api monitor by @mahiatlinux in #11340
Studio: pin the worker RSS in the two xet ledger tests that stand in for a worker by @danielhanchen in #11492
Give the GPU summary probe test the AMD helpers its extracted block calls by @danielhanchen in #11472
Studio: read the 16-bit pixels the way the rest of that test file already does by @danielhanchen in #11493
Studio: give back the continuation lease when the user stops the preflight by @danielhanchen in #9560
Studio installers: add --no-rollback, and report a full disk as a full disk by @danielhanchen in #11364
Studio: spell the owner's sidebar name Unsloth, not unsloth by @shimmyshimmer in #11275
fix(studio): bundle hf-xet for Desktop large Hub downloads by @Souravrajvi0 in #10849
Studio: anchor the worker source read on the test file, not the working directory by @danielhanchen in #11495
Studio: make the contrast slider reach surfaces and controls by @shimmyshimmer in #11459
Studio: scale spacing with the UI font size by @shimmyshimmer in #11458
Studio: make the drag preferences test compile, and ask the stricter question by @danielhanchen in #11503
Stop unsloth chat and unsloth inference ignoring the model's recommended settings by @NilayYadav in #11478
Studio: keep text visible in the reply when Thinking is off by @NilayYadav in #11481
Studio: keep a streaming code fence highlighted, and window its spans by @danielhanchen in #11423
fix(studio): track the repository resolved by the training loader by @Imagineer99 in #11462
Studio: promote -ts into tensor_split before manual strip by @indrajeetapache in #11334
studio/frontend: keep shutdown dialog open when server-stop fails by @danielhanchen in #5637
Repair a quarantined llama.cpp runtime at launch instead of failing at model load by @danielhanchen in #10494
Stop the installers importing the download stack just to be imported by @danielhanchen in #11446
Studio: show what the caches cost and let them be cleared by @danielhanchen in #10528
Stop unsloth chat and unsloth inference reloading a 16-bit model in 4-bit by @NilayYadav in #11477
Unsloth Studio: make a newly created API key selectable when the clipboard is unavailable by @LeoBorcherding in #11489
Cut CI queue: scope push triggers, and rebalance Repo tests (CPU) on measured runner time by @danielhanchen in #11504
Fix the update card clipping its own buttons at the 20px interface size by @danielhanchen in #11510
Studio: read the scaled lengths and colours the UI contracts were written against by @danielhanchen in #11505
Keep credentials out of cached directories, and audit the org against the miri cache disclosure by @danielhanchen in #11517
Stop dividing the GRPO eval loss by the accumulation steps by @vineethsaivs in #11337
studio: revamp the logs tab by @mahiatlinux in #11508
Studio: resolve upstream llama.cpp releases, assets and digests without the GitHub API by @danielhanchen in #5886
Build and publish signed prebuilt CUDA 13 wheels for torch 2.13 and 2.14 by @danielhanchen in #11515
Let the cache row's path label follow the interface font size by @danielhanchen in #11520
Give the Docker step harnesses the environment those steps now declare by @danielhanchen in #11521
Skip TRL's redundant kbit prep for models Unsloth already prepared by @danielhanchen in #11494
Keep generated RL configs on the reentrant gradient checkpoint path by @danielhanchen in #11519
Install the import-time Python dependencies before smoke testing a prebuilt wheel by @danielhanchen in #11527
Do not auto-enable padding-free for a model whose forward cannot take it by @danielhanchen in #11468
Find a GGUF companion published after the weights on the chat load route by @danielhanchen in #11448
Report a Linux ROCm APU host-backed pool of zero as measured, not unknown by @danielhanchen in #11451
Studio: validate the reasoning controls instead of coercing them by @Apoze in #9595
Studio tool loop: close two gaps in the static screens that gate model-written code by @danielhanchen in #11399
Fixes to XPU offloading during training by @ChowChowSonic in #9534
Studio: chat with a full fine-tune in 16-bit instead of a 4-bit copy by @NilayYadav in #11479
Docker: pass JUPYTER_PORT and the notebook skip options through run.sh by @NilayYadav in #11484
Let a pre-quant checkpoint be safetensors instead of a pickle by @danielhanchen in #11394
Rework Studio dark mode onto a single surface and text scale by @shimmyshimmer in #11502
Let a hosted pre-cast text encoder be safetensors, not only a pickle by @danielhanchen in #11454
Studio: warn when the disk is running low, and point at the caches by @danielhanchen in #11439
Add the Qwen-Image-2.1 family, and the Diffusers it needs by default by @danielhanchen in #11405
Route Qwen-Image-2.1 GGUFs to the sd.cpp engine by @danielhanchen in #11507
Resolve the auto class for omni checkpoints, and guard embedding lookups by @danielhanchen in #11467
Studio: stop one audio chat from blocking every other chat by @NilayYadav in #11483
fix(grpo): autocast with DEVICE_TYPE_TORCH instead of a pinned "cuda" by @li-lizhe in #11461
Warn when the installed transformers drops a pre-quantized checkpoint's bnb quant state by @danielhanchen in #11450
Install a prebuilt wheel's declared runtime dependencies before smoke importing it by @danielhanchen in #11536
Studio: replace reasoning pagination with a continuous transcript by @wasimysaid in #11522
Keep the compiler on for native architectures loaded with trust_remote_code = True by @danielhanchen in #11516
Studio: load a 16-bit Continued Pretraining run in 16-bit in Chat by @NilayYadav in #11480
Load prequantized multimodal checkpoints correctly on transformers 5.4 and 5.5 by @danielhanchen in #11452
Make main formatter-clean so pre-commit stops rewriting every PR by @danielhanchen in #11540
Export into a directory the temp-file test owns by @danielhanchen in #11532
Read the packing blocker's branch by behaviour, not by its spelling by @danielhanchen in #11534
Make the GPU probe harness name the helper it failed to extract by @danielhanchen in #11538
Make the thinking trace read as a trace, not as the answer by @shimmyshimmer in #11541
Load the published Qwen-Image-2.1 INT8 checkpoint, and four more follow-ups by @danielhanchen in #11535
Studio: take the hosted pre-cast fp8 text encoder by default on Qwen-Image-2.1 by @danielhanchen in #11539
Pin the Dynamic 2.0 4-bit text encoder for Qwen-Image-2.1's native route by @danielhanchen in #11542
Avoid 8-bit training optimizer on XPU by @ousamabenyounes in #10213
Lift the Classic light muted token to clear AA by @shimmyshimmer in #11544
Fix push_to_hub_merged for full fine-tunes on transformers 5 by @NilayYadav in #11482
Studio: reuse one database connection per thread for durable chat generation runs by @danielhanchen in #11525
feat(chat): make agentic turns survive tab-close by @CelesteHeartsong in #10365
Qwen-Image-2.1 fails every render at 20+ steps: FBCache cannot run on a prefix KV cache by @danielhanchen in #11546
Unsloth Studio: give the memory planner the companion split on a pipeline image load by @LeoBorcherding in #11548

New Contributors

@AzizMuminov made their first contribution in #10802
@taking-lying-flat made their first contribution in #11266
@TarzanZhao made their first contribution in #11238
@ChowChowSonic made their first contribution in #9534

Full Changelog: v0.1.811-beta...v0.1.812-beta

unslothai/unsloth v0.1.812-beta Qwen-Image-2.1 + Skills on GitHub