unslothai/unsloth v0.1.810-beta on GitHub

We're releasing our new updated Docker image along with multi-user accounts, RDNA1+2, FP8/INT8 diffusion support, ARM64 CUDA Windows support and many training, GRPO and inference improvements

Highlights:

New Docker with NVIDIA & AMD support: Guide
Multi user accounts with isolation Settings > Accounts
INT8/FP8 Image Diffusion inference - 2x faster
ARM64 Windows CUDA support for training, inference
GRPO improvements - Qwen3.5 support + lastest TRL, vLLM support
RDNA1, RDNA2 support for AMD
Better Windows NVIDIA GPU detection and recovery

Updated Docker image

Updated the main CUDA image and Studio setup, removed bundled caches, and restored Unsloth's training patches on GPU hosts.
Studio data persists on a volume without pinning application code. Updates, generated passwords, configurable ports, and package preservation are improved.
Added an AMD ROCm image alongside CUDA for supported Linux hosts.

Multi-user accounts

Create accounts from Settings > Accounts, with one-time setup codes and individual passwords.
Keep each account's work separate while sharing a loaded model when settings match. Existing single-account behavior stays unchanged.

Chat + reasoning

Edit, reorder, and steer queued prompts, including while a local model loads.
Added GGUF reasoning budgets, configurable chat width, desktop interface scaling, and a more compact composer.
Improved long-reasoning responsiveness, chat-history preservation, and handling of tool-generated files and images.

Hardware + inference

MLX gains video input, optional MoE and decode optimizations, and more reliable multimodal chats.
Improved GGUF memory estimates, DGX Spark handling, NVIDIA detection, NVLink checks, and user-selected GPU ordering.
Added native Windows ARM64 desktop packaging, Qwen3.5 fast_inference, and Ascend NPU device detection.

Research + API

Deep Research better preserves unfinished reports, handles JSON output, and reports runs without evidence. Scanned PDF uploads gain a local OCR fallback.
Improved provider reasoning controls, MCP compatibility, tool calling, and streamed error reporting.
Added video input on /v1/chat/completions and access to installed Ollama models through the API.

Training + installation

Added Data Recipes dataset downloads and fixed 16-bit fine-tune exports, LoRA saves, and Push to Hub.
Improved dataset handling, training memory estimates, and checkpoint saving on container shutdown. Core fixes cover normalization, RoPE, Q-GaLore, and distributed device selection.
More reliable Windows installs, GPU package repair, dependency caching, and verified offline installs.

Security + tool execution

Strengthened credential protection and tool-call validation, with approval requests for file access outside the sandbox.
Added preview OS-level tool sandboxing on supported Linux and macOS hosts. Network access remains unrestricted, and automatic mode retains existing safeguards when OS isolation is unavailable.

What's Changed

Bump install.sh / install.ps1 pin to unsloth>=2026.9.4 by @danielhanchen in #10640
Stop the uninstaller stranding old installs, and deleting directories that are not ours by @danielhanchen in #10471
Studio: correct the keyless tunnel copy, which promised more than the code does by @danielhanchen in #9690
fix(studio): keep GGUF compaction for multimodal chats by @Biotrioo in #10158
Fix Deep Research JSON output on MLX and Transformers by @Imagineer99 in #10611
Fold duplicated backend, hub and package tests into parametrized runs by @danielhanchen in #10499
Studio: reuse the named CLI API key across studio run by @claxman in #10608
Unbreak main: compare the autoload guard's conditions, not its exact text by @danielhanchen in #10654
Fix scanned PDF uploads with local OCR fallback by @Imagineer99 in #10620
Studio: size a local model by one copy of its weights by @NilayYadav in #10558
Studio: keep the LoRA rank, alpha and variant when leaving CPT by @NilayYadav in #10555
Studio: stop a generation recovery deleting the turn's tool calls by @shimmyshimmer in #10584
Studio: only strip a valid trailing IMAGES/RAG_SOURCES envelope by @NilayYadav in #10668
Read the installers' own code, not the scripts they emit by @danielhanchen in #10701
Raise the startup bundle budget to where main actually is by @danielhanchen in #10702
Studio: gate GGML_CUDA_P2P on a confirmed NVLink fabric, not a product name by @danielhanchen in #10646
Studio: stop pricing an Anthropic image by its base64 length in KV admission by @NilayYadav in #10669
Block markerless execution-class tool-call promotion in Studio by @danielhanchen in #10507
Studio: fail a Deep Research run when no step gathered any evidence by @NilayYadav in #10663
Studio: keep literal text visible when the request turns thinking off by @NilayYadav in #10662
Studio: keep an uploaded document's own name instead of replacing it with underscores by @NilayYadav in #10667
CI: trigger Studio smokes on what they observe, run install matrices as a PR subset, fold Chat UI shards by @danielhanchen in #10689
Windows No Compiler CI: see the intermediates the compiler cleaned up by @danielhanchen in #10711
Studio: keep the output a timed-out python or terminal call already produced by @NilayYadav in #10664
Studio: fit the composer on narrow screens, and scale the greeting with it by @shimmyshimmer in #10694
Studio: recover the GPU capacity nvidia-smi will not report on a DGX Spark by @danielhanchen in #10703
fix: materialize strided gradients before RMSNorm backward by @MrCapricornLiu in #10617
fix: exclude padded logits from softcapped cross entropy by @MrCapricornLiu in #10616
Studio: keep a chat working after switching to a model without tool support by @NilayYadav in #10454
Reduce comment volume across studio/backend/core by @danielhanchen in #10502
Reduce comment volume across the studio frontend and backend routes by @danielhanchen in #10503
Reduce comment volume across the installers, CLI, unsloth package and backend utilities by @danielhanchen in #10504
Cover the restore of every variable the install.ps1 handoff block saves by @danielhanchen in #10566
Studio: stop refusing a diffusion load on a DGX Spark against its own download by @danielhanchen in #10705
Refuse to format when ruff cannot run, and stop the fixed-point guard skipping itself away in CI by @danielhanchen in #10726
Let a Deep Research handoff survive a preamble beside the tool call by @BluePlexus in #10677
Keep the surrounding {{ }} when stripping bos_token from a chat template by @vineethsaivs in #10682
Studio: honor the vision run's gradient checkpointing setting by @NilayYadav in #10666
Send the thinking toggle to a "custom" provider by @BluePlexus in #10678
MLX CI: give the llama-server health wait the same 180s budget as the other health waits by @danielhanchen in #10674
Unbreak main: raise the startup transfer budget to the margin it shipped with by @danielhanchen in #10700
Upload merged models to the requested branch or pull request by @NilayYadav in #10760
Studio: keep long reasoning responsive with bounded pages by @wasimysaid in #10717
Studio: narrower desktop window and a sidebar edge trigger by @shimmyshimmer in #10706
Studio: Show Anthropic web search errors by @NilayYadav in #10757
Studio: Keep images returned by tools by @NilayYadav in #10762
Studio: Show an error when chat history cannot be restored by @NilayYadav in #10761
fix(dataprep): don't append EOS to full-size mid-stride chunks by @caiotheodoro in #10734
Windows installer: fix installs when the user profile path contains a space by @oobabooga in #10765
Stop capping unsloth chat at 512 tokens and the transformers/MLX paths at 2048 by @NilayYadav in #10665
Reject conflicting model loading settings by @NilayYadav in #10756
Docker Studio: stop disabling Unsloth's TRL patches on GPU hosts by @oobabooga in #10825
Unbreak main, and fix the five causes reddening the PR backlog by @danielhanchen in #10832
Stop a comment in unsloth_zoo from silently disabling the patched SFT trainer by @danielhanchen in #10854
Ask the vllm-free import tests in a clean interpreter, not in sys.modules by @danielhanchen in #10855
Re-format the two lines that drifted, so pre-commit stops failing on main by @danielhanchen in #10715
Studio: per-account isolation for shared installs and the desktop app by @danielhanchen in #10588
Use the short RoPE factor at exactly original_max_position_embeddings by @vineethsaivs in #10681
fix: do not default-on tool-call nudging on the external loop by @apurv-1 in #10685
Track vLLM tags through 0.29, and pin the bitsandbytes symbols by @danielhanchen in #10858
Give Repo tests (CPU) a budget it fits in, and name the patcher when it falls back by @danielhanchen in #10862
Studio: keep a stopped prompt in later chat turns by @claxman in #10625
Studio: accept zero-based shard numbering in cached model weights by @JohannesGezachew in #10857
Stop the CUDA spoof from deciding whether a GPU test skips by @danielhanchen in #10864
Studio: read the NVLink fabric from NVML instead of a 1.2s nvidia-smi shell-out by @danielhanchen in #10720
Studio: stop the log redactor backtracking quadratically on a cut ANSI sequence by @danielhanchen in #10721
Ask the spoof-survival question in a subprocess, not in the worker by @danielhanchen in #10866
Studio: prefill MLX VLM prompts at mlx-vlm's default step, fix cold-path position ids by @Lyxot in #10778
Pair the linearity measurements, so a busy runner stops reading as quadratic by @danielhanchen in #10868
Studio: add a Download all logs button to Settings > Logs by @danielhanchen in #10719
Speed up the three slowest tests without changing what they check by @danielhanchen in #10867
Clear the metadata directory cache before scanning installed records by @xy200303 in #10797
Studio: Keep API keys from training with the server's Hugging Face login by @NilayYadav in #10809
Studio: Export a full fine-tune as a real 16-bit model by @NilayYadav in #10808
Studio: Show an error when Anthropic fails a reply partway through by @NilayYadav in #10811
Studio: Let the agent run a command again after it edits a file by @NilayYadav in #10810
Studio: Stop at the requested stop text on transformers models by @NilayYadav in #10812
Stop clean_text deleting every non-ASCII character by @vineethsaivs in #10741
Let a duplicate tool call re-run after an intervening call (#10792) by @deepspace28 in #10846
fix(studio): avoid resident GGUF mappings for Windows no-reserve loads by @Imagineer99 in #10618
Studio: Estimate training memory correctly so the chat model can stay loaded by @NilayYadav in #10815
Name the method on the uploaded model card by @vineethsaivs in #10816
Replay tool-call arguments in the order the model generated (#10791) by @deepspace28 in #10845
Fix fp8 block dequant fallback on pre-sm89 GPUs by @xy200303 in #10790
Studio: Stop Deep Research deleting links from code in its report by @NilayYadav in #10814
Fix Q-GaLore quantization of single-sign groups by @vineethsaivs in #10841
Fix RAG dense retrieval on older SQLite runtimes by @alkinun in #10861
Docker: drop the baked uv cache from the Studio image and correct its labels and Hub page by @oobabooga in #10827
installer: the adaptive uv cache selection was unreachable on every writable install by @danielhanchen in #10626
Studio: Stop training the word None into CSV rows with blank cells by @NilayYadav in #10813
Skip the backgrounded-raw-bind prompt test on Windows by @danielhanchen in #10885
Fail closed when Studio exposure prompt is aborted by @wasimysaid in #10638
Docker Studio: keep Studio's data on a volume without pinning its code by @oobabooga in #10829
Docker: stop JupyterLab from silently taking Studio's port, and correct three host-script messages by @oobabooga in #10828
Advance the Q-GaLore Adam step counter once per update by @vineethsaivs in #10842
Run the Q-GaLore tests from the normal pytest job, not a workflow of their own by @danielhanchen in #10898
Accept the model transformers 5.x passes to create_optimizer by @danielhanchen in #10887
Fix three CI failures that were not the PRs they failed on by @danielhanchen in #10896
Pass Q-GaLore optimizer options to bitsandbytes by name by @vineethsaivs in #10874
Split Repo tests (CPU) into three shards, and take the shell suite off its critical path by @danielhanchen in #10882
Stop three studio meta-tests re-cutting and re-walking the same sources by @danielhanchen in #10886
Give every pwsh call in the suite the private startup cache by @danielhanchen in #10900
Share one child interpreter, and one notebook prototype, where the work is identical by @danielhanchen in #10893
installer: stop condemning a warm uv cache for a directory uv never writes by @danielhanchen in #10899
perf(studio): integrate optional MLX MoE and decode optimizations by @Lyxot in #10733
Studio: skip dependency-pass steps whose evidence holds, and keep that evidence on Windows by @danielhanchen in #10649
Allow fast_inference for Qwen3.5 by @danielhanchen in #10891
Studio: add Server default for vLLM Min P by @Etherll in #10623
Unbreak main: three files drifted out of the formatter, and one reads a file with the platform encoding by @danielhanchen in #10913
The install suite spends 1940s resolving the same few thousand paths over and over by @danielhanchen in #10916
Measure whether the runner label actually changes the queue wait by @danielhanchen in #10914
Studio: cap accelerate below 1.15 on Windows, where 1.15 breaks all ROCm training by @LeoBorcherding in #10819
Studio: read a video clip on MLX chat models by @Lyxot in #10480
Move the two highest-fire workflows off the ubuntu-latest label onto the same image by @danielhanchen in #10918
Add Agent Skills support to Studio by @wasimysaid in #10247
Studio: two edge failures the dependency-pass change left behind by @danielhanchen in #10915
Studio: keep a verified install when PyPI is unreachable and UV_OFFLINE is set by @danielhanchen in #10651
Studio: adapt macOS tray artwork to the menu bar appearance by @wasimysaid in #10880
Prevent agent startup from waiting on model catalog scans by @Imagineer99 in #10728
Studio: drop MediaPageLink tooltip below titlebar controls on Windows by @Souravrajvi0 in #10253
use the bundled icon first, and retry the cloudflared download by @NilayYadav in #10680
Docker Studio: keep unsloth-studio-update --ref from leaving Studio dead by @oobabooga in #10826
Studio: answer current llama.cpp, whisper.cpp and Node installs from their markers by @danielhanchen in #10648
Studio: do not lose a DGX Spark's memory pool to the page cache when fitting context by @danielhanchen in #10704
Studio: block sandbox bypass via command-position substitutions in terminal tool by @veriepicc in #10860
Honor user tensor_split when auto tensor-parallel planner returns an even split by @chakshu-dhannawat in #10884
fix(studio): share Run settings draft across sidebar and model picker by @Souravrajvi0 in #10851
Stop ten test assertions from measuring the runner instead of the code by @danielhanchen in #10897
Use the dynamic cache for Gemma multimodal generation by @danielhanchen in #10903
feat(docker): add an AMD ROCm image, for RDNA2 through RDNA4 and CDNA by @LeoBorcherding in #10820
Drop an MTP sidecar llama-server cannot load as --model-draft by @Lyxot in #10197
Studio: run the formatter on llama_cpp.py by @danielhanchen in #10926
Studio: Use saved Hugging Face login for model downloads by @NilayYadav in #10758
Shard the Python 3.13 backend pytest leg three ways by @danielhanchen in #10920
Run the unsloth_zoo suite beside Core instead of inside it by @danielhanchen in #10919
Release GGUF context while tool approval is pending by @Imagineer99 in #10673
Studio: add Download Dataset button for Data Recipes by @Souravrajvi0 in #10708
Studio: a CPU torch handover must not disarm the flavor invariant by @danielhanchen in #10906
Studio: Keep unfinished research reports by @NilayYadav in #10759
Studio: an AMD Windows host without usable ROCm takes Vulkan, not the CPU bundle by @danielhanchen in #10908
Studio: a masked NVIDIA GPU must not install the CPU bundle over CUDA by @danielhanchen in #10909
Studio: raise the micro-batch for the projectors that abort llama-server on an image by @chakshu-dhannawat in #10683
fix(chat_templates): remove CUDA hardcoded device from stopping criteria by @li-lizhe in #10684
feat(device_type): add Ascend NPU device detection by @li-lizhe in #10686
fix(sentence_transformer): resolve device consistently with sibling branches instead of hardcoding cuda by @li-lizhe in #10843
Studio: build the blocked-command alternation once instead of per call by @danielhanchen in #10927
fix(studio): stop MLX native controls from duplicating in streamed replies by @Lyxot in #10905
docker publish: drop the build caches by @danielhanchen in #10959
studio: compact the chat composer and add animated expansion by @mahiatlinux in #10863
Studio: Show local LM Studio, Ollama and Hermes models when running in Docker by @NilayYadav in #10936
Studio: keep the operator's package-manager policy in the installer by @danielhanchen in #10902
desktop: add interface scaling setting by @mahiatlinux in #9666
Studio: keep created files outside the collapsed tool card by @claxman in #10446
Tests: stop the live no-torch server test from uninstalling torch from the Studio venv by @oobabooga in #10955
Studio: make the model use the tool the client asks for by @NilayYadav in #10939
Studio: reinstall a missing PyTorch when unsloth studio update finds the package current by @oobabooga in #10956
studio: configurable HuggingFace endpoint via HF_ENDPOINT env var by @AdamPlatin123 in #4989
Studio: tell the model when a tool returns audio or a file by @NilayYadav in #10940
Studio: fix Push to Hub for base and quantized exports by @NilayYadav in #10943
Studio: keep earlier messages in chats with images by @NilayYadav in #10944
Studio: tell Codex when a chat is too long for the model by @NilayYadav in #10938
Studio: keep nested tool call fields llama.cpp drops when written out of order by @NilayYadav in #10935
Use model-neutral image generation tooltips by @shimmyshimmer in #10782
Allow removing missing images from the gallery by @shimmyshimmer in #10781
Fix inline Python plot rendering and downloads by @shimmyshimmer in #10954
Studio: remove New badges from Remote & LAN and Shortcuts by @shimmyshimmer in #10967
Studio: keep replaced replies out of ShareGPT exports by @NilayYadav in #10941
Studio: prevent chat options hover circle from overlapping titles by @shimmyshimmer in #10969
Studio: leave top_p out of external requests when set to Off by @NilayYadav in #10934
Studio: show the right reasoning effort and image options for every API model by @NilayYadav in #10933
Studio: make studio/setup.sh read the recorded uv cache, and probe the cache the CLI hands it by @danielhanchen in #10647
studio: centre the chat column by reserving the scrollbar gutter on both edges by @shimmyshimmer in #10966
Name the Ascend NPU as its own tests expect, and close the spoofable skip-guard hole by @danielhanchen in #10968
A first tool call must not answer from inside a legacy move's staging window by @danielhanchen in #10953
Take the wall clock out of two backend guards that fail on a loaded runner by @danielhanchen in #10952
Studio: add configurable chat width by @shimmyshimmer in #10958
fix: materialize strided LayerNorm inputs and gradients by @MrCapricornLiu in #10675
Studio: fix unsloth start when port 8888 is already in use by @NilayYadav in #10942
Studio: a Windows host whose NVIDIA GPU is hidden by CUDA_VISIBLE_DEVICES selects the CUDA bundle, not windows-cpu by @danielhanchen in #10976
Studio: accept the name field on chat messages by @NilayYadav in #10937
Keep a passthrough SSE stream alive while llama-server prefills by @danielhanchen in #10911
Studio: keep a working GPU llama.cpp prebuilt over a CPU-only source fallback, and say so in the footer by @danielhanchen in #10975
Studio: read the NVIDIA inventory from NVML or the CUDA driver when nvidia-smi cannot answer by @danielhanchen in #10985
Time the stall window test instead of counting polls by @danielhanchen in #10987
Say which unsloth_zoo revision Core tested, and warn when it is not main by @danielhanchen in #10988
Desktop: extend the NVIDIA X11 WebKit fence fix to native packages by @oobabooga in #10964
Run the chat width presets in CI, and two guards main is currently red on by @danielhanchen in #11001
Stop the video gate test failing on a rename it should not care about by @danielhanchen in #11003
Studio: expose image text encoder precision by @shimmyshimmer in #10788
Studio: show MTP companion download progress by @Imagineer99 in #10243
Studio: load the user's own Pi extensions and packages in unsloth start pi by @oobabooga in #10971
Close the partially initialized torch._dynamo window on the diffusion load path by @danielhanchen in #10981
Studio prebuilt selection: Windows CUDA lines ordered by detected DLLs rather than filtered, portable fallback attempt, keep paths require SM coverage by @danielhanchen in #11006
Stop committing the two fixture archives that antivirus vendors flag by @danielhanchen in #10979
Pre-import diffusers in the post-warm worker so the first image load does not by @danielhanchen in #10989
Studio: download image model assets without loading by @shimmyshimmer in #10789
fix(amd): turn on PyTorch's fast ROCm attention kernels on import unsloth, and drop the installers' host-wide export by @LeoBorcherding in #8821
tests: pin the amd/huggingfacenotorch extras and name the audit failure by @LeoBorcherding in #7536
Linux: index AMD visible-device masks by device, and floor gfx1102/RDNA4 to rocm6.4 by @LeoBorcherding in #9672
fix(AMD): recognise the RX 6950 XT, 6850M XT and 6550M instead of leaving them on CPU torch by @LeoBorcherding in #10746
fix(amd): let a pinned torch index carry a package extra, so RDNA1 can reach TheRock by @LeoBorcherding in #9138
Studio: Make installed Ollama models available through the API by @NilayYadav in #10763
Studio Linux update: torch repair and the source build read the NVIDIA driver library, a CPU wheel recorded as CUDA is repaired, a satisfied pin keeps the GPU prebuilt by @danielhanchen in #11008
Fix diffusion load cancellation during construction by @shimmyshimmer in #10780
Studio: stop iOS Safari zooming into chat text fields by @oobabooga in #10972
Capture which antivirus product is actually blocking installs, and stop discarding the false-positive submission packets by @danielhanchen in #10986
Format main to its own pinned hook, so pre-commit autofix can land again by @danielhanchen in #11019
Measure whether the host property can replace the native console thunk by @danielhanchen in #10984
Studio: stop a short chat video attachment from taking minutes to prefill by @oobabooga in #10973
Studio: ask before a chat tool reads or writes outside the sandbox by @danielhanchen in #10907
Studio installers read the NVIDIA driver library when nvidia-smi cannot answer, and setup.ps1 keeps a working GPU prebuilt over a source build by @danielhanchen in #11007
Studio runtime keeps the GPU when nvidia-smi cannot answer, and prefers a GPU-capable llama-server beside a CPU build by @danielhanchen in #11005
Give this work a before-and-after number, and find out the hardening already worked by @danielhanchen in #10996
fix(studio): keep Studio's auth directory out of the tool loop by @danielhanchen in #10895
WSL installer: stop running the C# compiler on the user's Windows machine by @danielhanchen in #10978
fix(studio): keep recipe workers alive when request threads retire by @alkinun in #11021
Delete the emit apparatus that existed to colour a banner, now that measurement says it was a no-op by @danielhanchen in #10999
Ask AMSI whether the installer compiles, which is the mechanism #10805 actually reports by @danielhanchen in #10994
Studio: organize Data settings and make archives searchable by @shimmyshimmer in #10974
Measure whether a hardening change alters what the installer does, instead of arguing that it does not by @danielhanchen in #10992
Studio: accept OpenAI video_url content parts on /v1/chat/completions by @Lyxot in #10439
Studio: add prompt queue controls and composer preferences by @shimmyshimmer in #10980
Studio: add reasoning budget controls by @alkinun in #7701
Studio backend tests: settle the real loggers package in conftest for repo-root pytest runs by @deepspace28 in #11028
fix(amd:audio): decode audio through PyAV when torchcodec installs but cannot load by @LeoBorcherding in #8642
Studio: re-measure the startup bundle budget, which main is red on by itself by @danielhanchen in #11013
Unsloth Studio Installer: improve GPU print details (NVIDIA, Intel, AMD) by @LeoBorcherding in #8762
Recognise a partially initialised torchvision and fix the studiobench ab_plan row by @danielhanchen in #11015
docs: clarify Docker shell and Windows setup instructions by @Imagineer99 in #11036
README: mount the Hugging Face cache in the Docker quickstart by @danielhanchen in #11062
Stop the deps CLI fixture racing another worker for the requirements tree by @danielhanchen in #11063
Release: sign every shipped PowerShell script and publish digests for the rest by @danielhanchen in #11069
Windows installer: stop uv splitting a requirements path, and keep the error position in diagnostics by @danielhanchen in #11064
Docker publish: verify the Studio image env, and stop telling latest users to opt in to CPU by @danielhanchen in #11071
Answer /status the runtime fields a backend double cannot invent by @danielhanchen in #11076
Give the launcher signal tests a worker to themselves by @danielhanchen in #11085
Re-draw a port the OS handed to another test by @danielhanchen in #11086
Assert the preset sheet subscribes, not how the subscription is spelled by @danielhanchen in #11065
Studio: keep waiting on a llama-server load that is still reading its model by @oobabooga in #11055
Studio: move composer settings into More by @shimmyshimmer in #11087
Studio: fix structured columns in Data Recipes on non-GGUF models by @NilayYadav in #11049
README: point AMD users at the ROCm image by @danielhanchen in #11066
docker: give unsloth/unsloth-rocm a Hub page and keep it in sync by @danielhanchen in #11059
Studio: on low disk, load the largest GGUF quant that fits and show which one loaded by @NilayYadav in #11052
Studio: organize chat settings and simplify descriptions by @shimmyshimmer in #11088
Studio: stop charging CPU-mapped input embeddings to the Metal context budget by @oobabooga in #11060
Studio: send a download over Xet when HTTPS cannot fetch its largest file by @oobabooga in #11061
tests: route the uv requirements-path pwsh call through the shared runner by @danielhanchen in #11091
Studio: soften authentication cards and update password setup copy by @shimmyshimmer in #11095
docker: correct four stale claims on the Hub page by @danielhanchen in #11099
Keep an abandoned zoom call out of the next test's stub by @danielhanchen in #11090
Studio: simplify and prioritize chat settings by @shimmyshimmer in #11093
Redact hf_token and wandb_token in unsloth train --dry-run output by @danielhanchen in #11100
Studio: stop context checkpoints from filling host RAM on hybrid recurrent models by @oobabooga in #11073
Studio: stop hiding MCP tools that have a dot or a long name by @NilayYadav in #11051
Studio: fix Base vs LoRA compare for messages with an image by @NilayYadav in #11047
Studio: keep the GPU order the user asked for instead of re-emitting it ascending by @LeoBorcherding in #11034
studio: preserve transcripts and recall audio and image models by @mahiatlinux in #10993
Windows on ARM: install the native ARM64 CUDA stack on NVIDIA hosts by @danielhanchen in #10282
docker: stop the toolkit installer reporting a healthy driver as missing by @danielhanchen in #11106
Studio: keep every turn when training a chat dataset with Alpaca format by @NilayYadav in #11053
Match stopping criteria separately for each generated sequence by @vineethsaivs in #11039
Apply Q-GaLore weight decay before the projected update by @vineethsaivs in #10931
Studio: make the GGUF memory estimate match the compute buffers llama.cpp allocates by @oobabooga in #11043
Studio: refine queue drag handles and resume icons by @shimmyshimmer in #11113
Fix the three Backend CI failures on main by @danielhanchen in #11111
Studio: stop merging the model twice when pushing a merged export to the Hub by @NilayYadav in #11050
docker: let the quickstart generate the passwords and print them by @danielhanchen in #11109
Studio: rewrite the guided tour and give every page one by @shimmyshimmer in #11105
Docker: let UNSLOTH_STUDIO_PORT change Studio's port inside the container by @NilayYadav in #11045
Keep an exported config's MTP declaration in agreement with the weights by @danielhanchen in #11042
Pin the win_arm64 scikit-learn to one version so the security audit cannot go red on its own by @danielhanchen in #11112
Windows on ARM: take the wheelhouse from our own signed release by @danielhanchen in #11108
Studio: clarify queue options and refine queue styling by @shimmyshimmer in #11114
Studio: save a training checkpoint when the container is stopped by @NilayYadav in #11044
Studio: make Thinking and repetition penalty work on messages with an image by @NilayYadav in #11048
Studio: let the GPUs picker set device order, not just membership by @LeoBorcherding in #11041
Studio: print the first-boot password when nothing else will show it by @danielhanchen in #11118
Studio: localize the prompt queue view and align its editor shortcut by @shimmyshimmer in #11117
Studio: repair CPU torch on an NVIDIA host from the Linux fast path by @danielhanchen in #11127
Studio: minimal OS sandbox for Python and Terminal tools on Linux and macOS by @danielhanchen in #10526
feat(studio): consistent SSH restrictions with approved-server allowlist by @Souravrajvi0 in #10642
Revert "feat(studio): consistent SSH restrictions with approved-server allowlist" by @danielhanchen in #11163
Repair the six CI checks that are red on main itself by @danielhanchen in #11139
Let the formatter fixed-point guard start on Windows by @danielhanchen in #11145
Read every uv and pip resolver flag through one boolish reader by @danielhanchen in #11138
Read the resident tensor names only where they are used by @danielhanchen in #11159
Stop two CI guards failing for the machine they ran on by @danielhanchen in #11164
Studio: keep the Hub token off redirect targets and clear a blank HF_ENDPOINT by @AdamPlatin123 in #11000
Studio: shrink the resume glyph, trim the queueing hint and round the scroll corner by @shimmyshimmer in #11179
Studio: report a model reply that finished with nothing in it by @shimmyshimmer in #11185
Stop a shared embedder setup stranding its model in the next test by @danielhanchen in #11178
Check the MLX audit cleaned up its own temp file, not the shared temp dir by @danielhanchen in #11181
Clear every GPU visibility mask the ROCm probe test drives production with by @danielhanchen in #11182
Run the formatter on the one file main left drifted by @danielhanchen in #11058
Studio: models.dev catalog for reasoning and image options on other API providers by @NilayYadav in #10957
Target unfused MoE expert parameters instead of dropping them by @danielhanchen in #11014
Studio: import ShareGPT chats with the right roles by @NilayYadav in #11149
Exit with code 130 when unsloth train is stopped with Ctrl+C by @NilayYadav in #11151
Studio: make UNSLOTH_DISABLE_UPDATE_CHECK also stop the llama.cpp and whisper.cpp checks by @NilayYadav in #11155
Studio: show disk space in the same GB as model sizes by @NilayYadav in #11150
Studio: pass search results and documents through to the model on /v1/messages by @NilayYadav in #11152
Keep the RoPE base frequency on transformers 5, and make triton optional by @danielhanchen in #11037
Studio: use --reasoning where available and name the compressed-tensors refusal by @danielhanchen in #11072
Resolve apply_qkv and apply_o defensively, and explain zero patch counts by @danielhanchen in #11038
Studio backend tests: pin the logical CPU count in the oversubscribed-threads spill test by @danielhanchen in #11120
Model picker: collapse a single downloaded quant again by @shimmyshimmer in #11190
Resolve the per layer device through a helper instead of a raw index by @danielhanchen in #11020
Installers: require x64 Python on Windows on ARM, and respect active conda by @danielhanchen in #11024
Report the video progress the GPU has actually done by @danielhanchen in #11004
Studio: reap the llama-server tree on Windows unload, and unblock a dead spawned backend by @danielhanchen in #11022
Pin every rank of a distributed launch to its own device, not just quantized ones by @danielhanchen in #11129
Keep every nested RoPE base, and never put a per-label mapping in rope_theta by @danielhanchen in #11192
Stop the empty-logits sentinel from claiming protocol dunders it cannot honour by @danielhanchen in #11130
Release: stop publishing standalone signed scripts by @wasimysaid in #11123
Studio: assert the argv, not the call count, in the shutdown spawn-refusal test by @danielhanchen in #10645
Studio: official image picks default to the hosted FP8 or INT8 checkpoint by @danielhanchen in #10883
perf(studio): keep MLX VLM generation on one stream by @Lyxot in #10962
Studio: serve every image an MLX conversation carries, not just the newest by @Lyxot in #10970
Studio: do not run Gemma tool call examples written inside code blocks by @NilayYadav in #11153
Move the Studio MLX pins to mlx 0.32.2 and mlx-vlm 0.7.1 by @Lyxot in #11080
Diagnose torch and triton version skew at import time by @danielhanchen in #11023
Take the compile-cache bundle write off the image generation request path by @danielhanchen in #11009
Fix the cold unsloth.save import, the preview role map and the GPU memory cap by @danielhanchen in #11077
Studio: keep a reasoning page start below its end by @danielhanchen in #10922
Define everything Refresh-Environment calls before the PSModulePath test runs it by @danielhanchen in #11194
Studio: serve an installed MLX checkpoint whose config drops its vision sub-config by @Lyxot in #10977
Stop a removed torchao class from killing plain LoRA by @danielhanchen in #11168
Fall back to HTTPS when hf_xet is installed but cannot be imported by @danielhanchen in #11167
Look for nvidia-smi everywhere it actually is by @danielhanchen in #11119
Turn the CUDA graph off for MiniMax-H3: measured 1.0018x for 3.93 GB by @danielhanchen in #11175
Unsloth Studio Installer: hold the install lock as a file as well as a mutex by @danielhanchen in #11094
Bump the actions group across 1 directory with 10 updates by @dependabot[bot] in #11199
Desktop: build and ship a native Windows ARM64 installer by @danielhanchen in #11122
Catch a temporary patch that stops applying instead of only logging it by @danielhanchen in #11198
Format the two files main is failing its own pre-commit hook on by @danielhanchen in #11203
Studio: honour an unmeasured context request and price the real attention plan by @danielhanchen in #11025
Route save_method="lora" to the adapter save instead of the merge by @danielhanchen in #11067
Installer: recover from and explain an unreadable llama.cpp cache on Windows by @danielhanchen in #10533
Studio: an AMD owner asking for a precision outright is told the real reason by @danielhanchen in #11032
Docker: keep the startup health checks off the HTTP proxy by @NilayYadav in #11157
Docker: stop notebooks from uninstalling the baked packages by @NilayYadav in #11156
Studio: stop declaring a live backend dead, and keep a crashed worker's stderr by @danielhanchen in #11070
Studio: judge a local chat model by the transformers 5 modality keys, not text_config by @danielhanchen in #11011
Baseline the four huggingface_hub 1.32.0 sandbox findings by @danielhanchen in #11206
Re-run the kwarg-spacing formatter on three files that landed unformatted by @danielhanchen in #11207
Kaggle GPU harness: an all-[N/A] listing is cannot attribute, not no GPU by @danielhanchen in #10472
Reject an hf_xet that is only an empty namespace package by @danielhanchen in #11205
Raise the unsloth_zoo floor to 2026.9.5 now that the compiled-path sentinel fix shipped by @danielhanchen in #11210

New Contributors

@MrCapricornLiu made their first contribution in #10617
@BluePlexus made their first contribution in #10677
@caiotheodoro made their first contribution in #10734
@apurv-1 made their first contribution in #10685
@JohannesGezachew made their first contribution in #10857
@xy200303 made their first contribution in #10797
@deepspace28 made their first contribution in #10846
@veriepicc made their first contribution in #10860
@li-lizhe made their first contribution in #10684

Full Changelog: v0.1.808-beta...v0.1.810-beta

unslothai/unsloth v0.1.810-beta Docker + Multi User + AMD Support on GitHub

Highlights:

Updated Docker image

Multi-user accounts

Chat + reasoning

Hardware + inference

Research + API

Training + installation

Security + tool execution

What's Changed

New Contributors

unslothai/unsloth v0.1.810-beta
Docker + Multi User + AMD Support

on GitHub