unslothai/unsloth v0.1.905-beta on GitHub

Turn any text or vision LLM into a Jev-style decision model in Unsloth, with decision accuracy going from 30% to 80%. Train, test, export and serve decision models directly from Unsloth. Also included: native ComfyUI models, diffusion improvements and a better Browser in Desktop.

Highlights

8th Oct Fixes - Better Browser + 100+ bug fixes + perf fixes
Sandboxing with Bwrap for Linux, Seatbelt for Mac and MXC for Windows
Turn any model into a Jev-style decision model. Accuracy went from 30% to 80%
Load ComfyUI diffusion models natively in Unsloth
Faster + more accurate diffusion with INT8 ConvRot and more

dt-demo.mp4

Decision models

Train any text or vision LLM as a Jev-style decision model using QLoRA.
Test trained models directly from the Decision API settings.
Make decisions with confidence scores for every option.
Export Clef models with Qwen3.5 backbones and Laya models to GGUF.
Serve supported decision models through llama.cpp, including models that understand images.
Save and resume smaller adapter and decision-head checkpoints.
Guide at https://unsloth.ai/docs/basics/train-your-own-decision-model-with-unsloth

Diffusion + ComfyUI

Run supported ComfyUI image and video models directly from Hugging Face.
Unsloth now recognises ComfyUI checkpoints and local model folders automatically.
Use your ComfyUI text encoders and VAEs in Unsloth.
Run Krea-2, HunyuanImage-2.1 and Wan2.2 expert pairs, plus ComfyUI NVFP4 and MXFP8 models.
Qwen-Image-2.1 now keeps more full-precision image detail with INT8 ConvRot enabled by default.
Faster Qwen-Image-2.1 ConvRot generation on supported NVIDIA GPUs.

Training + performance

Train Qwen3.5-35B-A3B up to 4.1x faster and Qwen3-30B-A3B up to 3.3x faster with QLoRA on A100 and RTX PRO 6000.
Improved sample packing for gated-delta, Mamba2 and short-convolution models.
Train prompt and completion message lists as one conversation.
Vision datasets now keep each row's own question.
Chat exports and training data now include the system prompt.

Browser + Desktop

Ask about open pages in the Desktop Browser.
Confirm Browser downloads and choose where files are saved.
Reorder pinned pages in the sidebar like chats.
Search continues past unusable results and can fall back to Wikipedia.
Reply citations such as [1] now open as links.
Pick, pin or unload RAG embedding models directly from the RAG menu.

Sandboxing

Bwrap on Linux, Seatbelt on Mac and MXC on Windows sandbox code the model runs.
View sandbox status and choose protection levels in Settings.

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

Platform	Link
Windows	Download
macOS	Download
Linux x64 / Ubuntu (deb)	Download
Linux ARM64 / Ubuntu 24.04+ (deb)	Download
Linux x64 (AppImage)	Download
Windows ARM64	Download

What's Changed

Bump install.sh / install.ps1 pins to unsloth>=2026.10.1, unsloth-zoo>=2026.10.1 by @danielhanchen in #12869
Studio: list embeddinggemma-2 first in the embedding model picker by @shimmyshimmer in #12870
Repair three checks that went red on main with the 10-06 Studio merges by @danielhanchen in #12868
Studio: run ComfyUI-format video quants on the int8 / fp8 runtimes by @danielhanchen in #12851
Studio: free PyAV's per-thread scalers before a fork so preexec_fn spawns still exec by @danielhanchen in #12863
Tests: give the setup.ps1 download progress pwsh its own startup cache by @danielhanchen in #12882
Studio: fill the Hebrew and Swedish strings that left the strict i18n check red on main by @danielhanchen in #12881
Baseline the eight unsloth-zoo 2026.10.1 findings after review by @danielhanchen in #12884
Support every PEFT init_lora_weights option, with fast PiSSA and MiCA init by @Suchitra-idu in #6879
Sandbox test: wait for the cache scan workers before counting them by @danielhanchen in #12896
Studio: keep browser panel tooltips, toasts and menus visible over desktop web pages by @oobabooga in #12895
Studio: continue searching past unusable results and add a Wikipedia fallback by @oobabooga in #12892
Studio: list every Transcribe ASR model in Voice settings by @Etherll in #12898
Frontend test: give the cold Vite SSR render in reasoning-source-render room on a loaded runner by @danielhanchen in #12903
Run shell suites and Windows browser checks in parallel by @oobabooga in #12899
Studio: add a New badge beside Audio in the sidebar by @Etherll in #12891
Composer settings driver: poll the submitted list instead of reading it once after the key press by @danielhanchen in #12920
Studio: support current native builds and preserve CPU asset selection by @oobabooga in #12902
Studio: pick the quant of a GGUF dictation model in Voice settings by @Etherll in #12900
fix(studio): stop a managed runtime when the client drops its stream by @goodmai in #12266
Studio: train prompt/completion message lists as one conversation by @NilayYadav in #12910
Studio: keep each row's own question when training on a vision dataset by @NilayYadav in #12909
Studio: train transparent PNG and WebP images on white instead of black by @NilayYadav in #12908
Use the requested max_seq_length for encoder embedding models by @NilayYadav in #12915
Studio: show the LAN address on the API page when LAN access is on by @NilayYadav in #12906
Studio: hide the negative prompt on image models that ignore it by @NilayYadav in #12914
Keep the notebook and saved model when unsloth-run runs a URL by @NilayYadav in #12907
Studio: drag pinned pages in the sidebar like chats by @shimmyshimmer in #12927
Studio: Ask about this page on every page, desktop included by @shimmyshimmer in #12926
Docker: allow unsloth_root_shim.py into the build context by @danielhanchen in #12929
Tauri transport test: start the late backend after the old ladder is spent, not at 3s by @danielhanchen in #12930
Studio: simpler icons for the audio pages, one audio icon in the Library by @Etherll in #12890
Desktop contract: count #12927's scaled sidebar row by @danielhanchen in #12931
fix(studio): preserve skill mention intent and denied preload context by @wasimysaid in #12841
Studio: embedding model picker, pins and eject in the RAG menu by @shimmyshimmer in #12875
install-kernels: skip mamba_ssm below sm80 by @danielhanchen in #12921
Gemma-4 26B/31B: train with the empty thought channel on non-thinking turns by @danielhanchen in #12867
Studio: try a decision from the Decision API settings by @NilayYadav in #12916
Studio: confirm browser downloads, choose the download folder by @shimmyshimmer in #12832
Studio: default Qwen-Image-2.1 int8 to the hosted ConvRot file, with shared rotations by @danielhanchen in #12874
Studio: recognise ComfyUI checkpoints by name and header, and list ComfyUI model folders by @danielhanchen in #12878
studio: install the gstreamer recording plugins with the deb package by @mahiatlinux in #12905
Studio: turn [1]-style citations in replies into links by @NilayYadav in #12912
Studio: include the chat's system prompt in exports and training data by @NilayYadav in #12913
studio: guard project submits during IME composition by @mahiatlinux in #12924
studio: scope speech download cancellation to its attempt by @mahiatlinux in #12925
studio: keep general settings from restoring stale tokens by @mahiatlinux in #12922
Studio: show when Windows MXC already runs in the built-in container by @danielhanchen in #12938
Studio: key the diffusion compile cache by the loaded quant variant by @danielhanchen in #12887
Studio: rebuild an image / video GGUF from the cached copy when only its header changed by @danielhanchen in #12928
Scope UNSLOTH_HIGH_PRECISION_LAYERNORM to the load that sets it by @danielhanchen in #12873
fix(studio): honor llama.cpp update dismissal and snooze by @wasimysaid in #12934
Keep flash attention from reading Qwen3.5 mRoPE position ids as packed sequences by @danielhanchen in #12856
Studio: load Wan2.2-A14B expert pairs and tell LTX-2.3 distilled from dev by its weights by @danielhanchen in #12872
Train a decision model from a plain language model by @danielhanchen in #12772
Train any text or vision LLM as a Clef decision model from Studio, with FastDecisionModel.predict and adapter saves by @danielhanchen in #12876
Studio: install transformers releases needing hub >= 1.31, and transformers main after consent by @danielhanchen in #12871
Correct sample packing for hybrid models (gated-delta, Mamba2, short conv) by @kfastino in #9812
Studio: read replies aloud without markdown symbols by @AzizMuminov in #12598
Studio: finish an update with the setup script it installed by @Etherll in #12897
Serve decision models through llama.cpp in Studio, and export them to GGUF by @danielhanchen in #12939
studio: fix live monitor background in light mode by @mahiatlinux in #12904
Studio: show the hosted text encoder download in image load progress by @Etherll in #12894
Studio: update the audio.cpp runtime from the in-app update by @Etherll in #12893
Settings contract: read the embedding picker's stacking classes inside cn() too by @danielhanchen in #12945
install-kernels: install mamba_ssm on sm75 with Triton 3.4+ by @danielhanchen in #12944
Studio: load Wan2.2 hosted FP8 / INT8 files, and hosted files under low_vram by @danielhanchen in #12888
Studio: load a single .safetensors DiT from any Hugging Face repo by @danielhanchen in #12879
fix(studio): remember per-GPU layer ratios by @Imagineer99 in #12774
Studio: load ComfyUI Krea-2 and HunyuanImage-2.1 single-file DiTs by @danielhanchen in #12885
Studio: load ComfyUI text encoder and VAE files beside a single-file DiT by @danielhanchen in #12883
Studio: load ComfyUI nvfp4 and mxfp8 DiT single files by @danielhanchen in #12877
Studio: keep $PATH, $HOME and other shell variables as text, not maths by @NilayYadav in #12911
Studio: price the context meter off a tool loop's final pass, not the whole turn's completions by @sumingwang233 in #12889
Studio: open Try a decision with the trained model after Use in Decision API by @NilayYadav in #12953
Studio: harden browser downloads after #12832 by @danielhanchen in #12943
Repair four checks that went red on main with the decision-model merges by @danielhanchen in #12954
Bump install.sh / install.ps1 pins to unsloth>=2026.10.2 by @danielhanchen in #12960
install-kernels: tidy the sm75 mamba_ssm follow-ups by @danielhanchen in #12948
Studio frontend: bump proxy-addr, seroval and MCP SDK for npm advisories by @danielhanchen in #12932
Studio: close managed-account and API-key gaps in owner-only routes by @danielhanchen in #12940
Studio: harden S3 dataset keys, uv fallback, header reads and auth body cap by @danielhanchen in #12936
Decision models: follow-up fixes after #12772, #12876, #12939 by @danielhanchen in #12949
Fix the backend CI guards main fails after the decision and ComfyUI merges by @danielhanchen in #12958
Baseline the two unsloth-zoo 2026.10.2 findings after review by @danielhanchen in #12956
Installer differential: ignore winget spinner frames in the transcript by @danielhanchen in #12965
tests: keep the Kaggle launcher's signal handlers out of the pytest worker by @danielhanchen in #12974
Stop test_dataset_cache_safe leaking the Hub no-symlink switch into later tests by @danielhanchen in #12973
Studio: keep Deep Research tables intact when a cited title has a pipe by @NilayYadav in #12990
Studio: keep tool call arguments in Qwen3.5 safetensors and MLX prompts by @NilayYadav in #12988
Fix linked-folder indexing of hidden subdirectories by @Imagineer99 in #12972
Studio: compact long chats on self-hosted connections to the window the server reports by @oobabooga in #12975
Studio: show project sources as unused on models without tools by @NilayYadav in #12993
Studio: show a download card when the python tool edits an attached file by @NilayYadav in #12992
Studio: stop crashing on a lowercase boolean in PYTORCH_ALLOC_CONF by @oobabooga in #12976
Studio: decode pages in the browser panel the way browsers do by @NilayYadav in #12983
Studio: use llama-server for embedding models the installed sentence-transformers cannot load by @oobabooga in #13005
Studio: train vision datasets that have some rows without an image by @NilayYadav in #12991
Support text attachments in per-chat prompt queues by @Imagineer99 in #12964
Studio: make Thinking off and Preserve thinking work on Qwen3.6 by @NilayYadav in #12989
Studio: smaller settings info icons, engine notes under the engine name by @shimmyshimmer in #13008
Studio: static dark dropdown glow, and stop modal opens restyling the page by @shimmyshimmer in #13013
Studio: open video attachments in the browser panel by @shimmyshimmer in #13007
Studio: show a site's own error page in the browser panel by @NilayYadav in #12986
Studio: list typed-decisions datasets first for decision training by @NilayYadav in #12982
Keep the model a training script saves when it runs in Docker by @NilayYadav in #12994
Skip the fast LoRA paths when lora_B has a bias by @vineethsaivs in #12981
Studio: prefer llama-server for EmbeddingGemma in the embedding model picker by @oobabooga in #13006
Studio: ask before numpy pickle loads, keep audio tags inline, cap the variable-prose regex by @danielhanchen in #13001
Studio: parse Yahoo's newer result layout in web search by @oobabooga in #12980
Update the flex large head dim mask tests to the unpadded causal-mask skip by @danielhanchen in #13018
Studio: serve whisper-server under a random per-launch request path by @danielhanchen in #13002
Studio: find nvidia-smi under WSL on every query, count Core Ultra Arc iGPUs as XPU, cut GPU masks at an invalid index by @danielhanchen in #12962
Studio: make llama-fit-params executable after the macOS prebuilt install by @jayzhou2309 in #12917
Studio: leave a bracketed IPv6 host unchanged in dial_host by @drakeo338 in #12733
Studio: build VAE tile blend weights on CPU so MPS tiled encode works by @gokay-ai in #12937
Studio: run llama-server with a per-launch API key by default by @danielhanchen in #13011
Studio: keep new image sets from merging into an existing one by @NilayYadav in #12987
install.sh: install for the discrete AMD GPU when an iGPU is listed first by @danielhanchen in #12963
FastModel: fall back to Unsloth inference on GPUs older than Volta by @danielhanchen in #12959
Unsloth Studio (AMD): show each GPU's live VRAM and utilization under its own HIP id by @danielhanchen in #12961
Studio: Downloads button in the browser panel by @shimmyshimmer in #13009
Studio: keep a chat's HTML pages with that chat by @NilayYadav in #12985
Studio: make the browser's right-click downloads work on macOS by @shimmyshimmer in #13003
Studio: put the skill row chevron next to the skill name by @shimmyshimmer in #13033
Studio: close refused tool calls under tool_choice none by @Beverly621 in #12627
Studio: keep browsing from a temporary chat out of browser history by @NilayYadav in #12984
Studio: duplicate a past training run into a new Configure draft by @Padi142 in #12979
Unsloth Studio / Desktop: show a cached image GGUF as Partial until Run has its text encoder and VAE by @LeoBorcherding in #12557
Load one tensor at a time when quantizing a 16-bit checkpoint, so Qwen3.5-27B 4-bit loads without Block Swap by @LeoBorcherding in #12997
Studio: show the real cause when a desktop install runs out of disk space by @huntersgordon in #12919
Fix notebook failures on Kaggle T4x2 and duplicate import warnings by @danielhanchen in #13024
Studio: skip the Xet probe's GPU-init-off zoo retry on GPU hosts by @arcusbuilds in #12478
Studio: stop MXC read grants looping on a Microsoft Store Python by @danielhanchen in #13025
Studio: avoid a TypeError after a tools module reload by @jiangLLM in #11506
Studio: retry web search over HTTP/1.1 when the connection is reset by @danielhanchen in #13032
Studio: stop a Git Bash nul file from blocking every Windows MXC tool call by @danielhanchen in #13037
Decline the packed INT4 kernel when weight_scale has the wrong group count by @jayzhou2309 in #12957
Studio: let another account join the resident GGUF instead of replacing it by @danielhanchen in #13038
Studio: give the annotate comment box a visible shadow in dark mode by @shimmyshimmer in #13071
Installer: read the venv's torch in isolation so a PYTHONPATH torch cannot break the install by @danielhanchen in #13041
Studio: keep tensor split and MTP when Auto would pick a DFlash drafter that aborts it by @danielhanchen in #13040
Allow backward through eval-mode and for_inference forwards by @danielhanchen in #13050
Make Q-GaLore optimizer state resumable from checkpoints by @danielhanchen in #13051
Studio: Docs links for Sandbox, Agents and the Decision API by @shimmyshimmer in #13076
Studio: serve decision models through the MLX engine on Apple Silicon by @Lyxot in #13015
Studio: pin FastFlowLM 1.0.7 for Qwen3.8 27B on the AMD NPU and keep Lemonade / FastFlowLM current by @danielhanchen in #13047
Studio: report the real llama.cpp version for source builds by @danielhanchen in #13036
Studio: name media companion downloads by component, and show what Run still fetches for an on-device GGUF by @danielhanchen in #13027
Studio: keep a dragged annotate area the size it was drawn by @shimmyshimmer in #13070
Studio: load a repo id from its scan-folder copy instead of re-downloading it by @danielhanchen in #13034
Train EXAONE 3.5: name the token embedding remote code no longer exposes by @danielhanchen in #13058

New Contributors

@kfastino made their first contribution in #9812
@sumingwang233 made their first contribution in #12889
@drakeo338 made their first contribution in #12733
@Beverly621 made their first contribution in #12627
@Padi142 made their first contribution in #12979
@huntersgordon made their first contribution in #12919
@arcusbuilds made their first contribution in #12478
@jiangLLM made their first contribution in #11506

Full Changelog: v0.1.903-beta...v0.1.905-beta

unslothai/unsloth v0.1.905-beta Train your own Decision model on GitHub