unslothai/unsloth v0.1.806-beta on GitHub

Run Qwen3.8-Flash-Next and GLM-5.3-Flash up to 2x faster with MTP. MTP is enabled by default, you can still disable it.
Also our new release includes 170+ training, chat, hardware, and performance improvements.

Highlights

Smoother model loading (less errors) across local servers and connected providers.
Safer chat edits that preserve tool cards, reply details, and conversation branches.
New local media APIs for video, audio, and MLX-served models.
New audio support with new models, progress tracking including: MiniMax-Music3, Higgs, MOSS and more!
Improved multi-GPU planning, memory fitting, and split-model training.
Strengthened AMD/ROCm detection, installation, and GPU compatibility.
Upgraded MCP, Deep Research, OAuth, and agent tool reliability.

Qwen3.8-Flash + GLM-5.3-Flash

Qwen and GLM now generate faster with MTP enabled by default.
Use GLM tools across longer, multi-turn chats.
Qwen automatically applies the recommended settings for thinking and non-thinking modes.

Download Qwen3.8-Flash-Next and GLM-5.3-Flash. See the Qwen guide and GLM guide for recommended settings and available GGUFs.

Faster MLX inference

Fine-tune both large MoE models with text or images on Apple Silicon using MLX.
Long Qwen chats now run much faster on Mac, with follow-up turns up to 30x faster.
MLX models now use their full context size and support much longer batched generation.
MLX releases GPU memory more cleanly between generation bursts and model switches.
Serve MLX models through Unsloth's OpenAI-compatible API.

Audio

Added support for MiniMax-Music3, Higgs, MOSS audio models.
Added live progress updates while audio is being generated.
Audio clips can now be archived and managed.
Improved reliability with custom TTS playback fixes, Whisper pairing checks, and stronger audio testing.

Chat + tools

Run several tool calls at once without mixing up their arguments.
Keep tools available when chatting with images.
Each chat keeps its MCP connection for faster tool calls.
Local models can edit code using Codex’s apply_patch tool.
Continue long chats with images and other media using Auto Compaction.
Review and approve Deep Research plans before research starts.

Training + hardware

Train larger models across multiple GPUs with automatic placement.
AMD installs choose the best build across Windows and Linux, with BF16 on more GPUs.
Export GLM-5.3 MLX fine-tunes to GGUF.
Choose custom GGUF shard sizes and save locations.

API + Desktop

Generate videos through the new OpenAI-compatible Videos API.
Updates download in the background and install when you restart.
Choose a custom port for LAN access.
Generate audio with Higgs, MOSS and MiniMax models.
Track audio generation progress and archive finished clips.
Model downloads show clearer progress and can switch from Xet to HTTP automatically.

Download Unsloth Desktop

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

Windows
macOS
Linux

🦥 Download Unsloth Desktop

What's Changed

Put the smart offload planner back behind its flag by @danielhanchen in #9862
Studio: stop the per-chunk autosave writing back messages the server owns by @danielhanchen in #9865
Bump install.sh / install.ps1 pins to unsloth>=2026.8.22 by @danielhanchen in #9868
Fix Studio hydrating synced GGUF files before selection by @milewski in #9539
Fix datasets PyArrow registration after failed warm import by @Imagineer99 in #9626
fix(studio): stop offering a transformers upgrade where it cannot load anything by @Lyxot in #9878
Fix Studio CPT overwriting LFM2 all-linear LoRA targets by @Souravrajvi0 in #9870
fix(studio): accept trailing slash for model discovery by @Imagineer99 in #9876
studio: accept shared youtube clipboard links by @mahiatlinux in #9903
studio: tell users when a canvas is blocked by the network access setting by @NilayYadav in #9894
studio: fix streamed tool calls in studio api monitor by @Etherll in #9896
studio: serve installed mlx models through the openai-compatible api by @mahiatlinux in #8768
Fix sidebar footer width and context menu radius by @shimmyshimmer in #9918
strip inherited claude provider routing from local agent launches by @Souravrajvi0 in #9871
studio: honor a cancelled load inside the llama-server health wait by @NilayYadav in #9893
fix model hub recent sorting for on-device models by @Imagineer99 in #9642
add a setting that tells the model the current date by @mahiatlinux in #8879
studio: fix lan address actions and launch bind reporting by @Lyxot in #9170
rag: avoid re-hashing linked-folder snapshots in start_ingestion by @DhruvaMyakeri in #9913
fix(studio): coalesce concurrent keyless-settings reads to stop SQLite thundering herd by @Greninja44 in #9911
Studio: add the OpenAI videos API on /v1/videos by @NilayYadav in #9891
Clarify tool approval copy by @Imagineer99 in #8738
Pin Claude Code to the loaded model's real context window by @NilayYadav in #9940
studio: drop the schema bounds llama.cpp's grammar engine cannot compile by @Lyxot in #9923
Fix shell selection for remote Agents commands by @Imagineer99 in #8844
Studio: accept the loadConfig keys a saved system-prompt preset already sends by @indrajeetapache in #9881
Studio: compact long chats with media attachments by @Imagineer99 in #9476
Studio: add Higgs, MOSS, and MiniMax audio models by @Etherll in #8794
studio: add ordered arguments to mcp server settings by @Etherll in #9943
studio: list route-aware models through the openai api by @NilayYadav in #9892
fix(studio): repair custom TTS endpoint test and playback by @Souravrajvi0 in #9905
Studio: render escaped inline math in lists by @Imagineer99 in #9796
deprecate unsupported custom prompt templates by @vineethsaivs in #8938
Restore UNSLOTH_RETURN_LOGITS after prediction_step instead of forcing it to 0 by @danielhanchen in #6239
Studio: update in the background and restart when ready by @NilayYadav in #9890
Serve every MiniMax H3 asset from unsloth rather than a community repack by @danielhanchen in #8832
fix(studio): gate external tool nudges and preserve retry context by @Biotrioo in #9125
Chat: let the LoRA compare rehydrate a pair the generalized compare saved by @lonexreb in #9908
fix: expose codex apply_patch for local models (#9114) by @Biotrioo in #9121
Fix duplicate GGUF rows in custom model folders by @Imagineer99 in #9975
Revert "fix(studio): gate external tool nudges and preserve retry context" by @danielhanchen in #9990
studio: isolate mcp browser test state by @mahiatlinux in #9951
Fix three CI regressions: code-span anchor, load-dispatch anchor, recordless pip install by @danielhanchen in #9959
studio: fix audio ci tests by @mahiatlinux in #9953
studio: fix openai route ci tests by @mahiatlinux in #9952
Skip the inline-code mask when every backtick is inside a fence by @danielhanchen in #9991
stop the planner withholding a split reserve from one card, and price physical cores by @danielhanchen in #9863
Parse the downloaded model mapper instead of exec'ing it by @danielhanchen in #9777
Recapture the control leg reference against the fixed loss normalisation by @danielhanchen in #9569
Stop os.dup2 alone being a reverse shell, and allowlist the zoo URL guard by @danielhanchen in #9994
Give the in-venv CLI tests a writable STUDIO_HOME by @danielhanchen in #9993
Settle the compact layout before measuring the permission pill by @danielhanchen in #9992
Fail CI on a new exec/eval/compile of a value that is not written out by @danielhanchen in #9999
Advertise a backend version a released AppImage will accept by @danielhanchen in #10001
Budget the dyld classifier in CPU time, not wall clock by @danielhanchen in #10003
Stop one slow SQLite writer from becoming a stream of Studio log errors by @danielhanchen in #9996
studio: cache the servability scan behind GET /v1/models by @danielhanchen in #9998
Make the multi-turn smoke actually assert that history reached the model by @danielhanchen in #10009
Share the fixture backend version with the model-download E2E by @danielhanchen in #10002
Give the cancel drain a budget sized for a loaded runner by @danielhanchen in #10008
Size the model picker's row metadata to the ink it holds by @shimmyshimmer in #10006
Studio: fix unscrollable model dropdowns on the export page by @Lyxot in #10020
Ground the multi-turn smoke on the turn that survives a 270M model by @danielhanchen in #10025
Studio: honor the effort level Claude Code sends on /v1/messages by @NilayYadav in #9941
Studio: archive audio clips by @NilayYadav in #9637
Pre-flight whisper pairing before the chained install by @danielhanchen in #10027
Dim an over budget model row instead of pilling every line by @shimmyshimmer in #10007
studio: reap durable chat generation runs that stop making progress by @danielhanchen in #9997
Stop three Backend CI tests measuring the runner instead of the code by @danielhanchen in #10035
Let the AppImage E2E use the Retry the failed quant row offers by @danielhanchen in #10032
Keep a modules_to_save copy on the card it was loaded onto by @danielhanchen in #10034
Train a model the planner split across cards, instead of dying in index_select by @danielhanchen in #9995
Name the encoding when the offload test reads llama.py by @danielhanchen in #10039
Answer llama-server discovery probes instead of serving them the app shell by @danielhanchen in #10037
Ask one formula whether a GGUF fits by @shimmyshimmer in #10029
Stop the llama-server probe deny-list answering 405 by @danielhanchen in #10049
Stop the subagent tool pre-approval swallowing the forwarded prompt by @NilayYadav in #9939
Use the plain download icon in the Hub On Device views by @shimmyshimmer in #10055
Tag the compat endpoint's Ollama rows with their real source by @lonexreb in #9988
Stop the GGUF header and the chat detail polls from filling the log by @danielhanchen in #10040
feat(export): GGUF shard size control, save directory override, fix private Hub push by @LeoBorcherding in #6107
Studio: set GGML_CUDA_ENABLE_UNIFIED_MEMORY only where it gains memory by @oobabooga in #9884
Wait for the idle batch flush instead of racing a fixed sleep by @danielhanchen in #10063
Use the head-aware planner for Studio's multi-GPU device map by @danielhanchen in #10042
Let tool chats decode in parallel instead of one at a time by @danielhanchen in #10046
Fix excessive studio.db writes during chat inference by @Imagineer99 in #9945
Stop shipping test suites in the wheel and the sdist by @danielhanchen in #10060
Studio: keep a Windows update from leaving a CPU-only PyTorch by @danielhanchen in #9857
Read the chat storage keys on call, not at module scope by @danielhanchen in #9844
Studio: install torch from an index that has kernels for the user's AMD GPU by @oobabooga in #9829
fix(studio): normalize Muse Glimmer's recipient-addressed protocol by @Lyxot in #8422
fix(studio): read tool-call arguments as the types their schema declares by @Lyxot in #10019
Studio: notice a quarantined payload before trusting the update fast path by @danielhanchen in #10053
Say why Studio could not enumerate a GPU, instead of just logging [] by @danielhanchen in #10056
CI: fix llama.cpp seed caching, Diffusers wheel installs, and isolated test coverage by @oobabooga in #9979
Studio: resume HTTP after Xet stall without a transport-conflict banner by @Souravrajvi0 in #10045
Fix GLM-5.3 chats breaking after a tool runs by @NilayYadav in #9938
Studio: fix clipped settings scrollbar on Windows WebView2 by @NilayYadav in #10074
Bump dompurify from 3.4.2 to 3.4.14 in /studio/frontend in the npm-frontend-security group across 1 directory by @dependabot[bot] in #6753
Studio: show progress during audio generation by @Etherll in #10081
Show activity on a 0-byte Xet download and keep finished cards visible longer by @NilayYadav in #9303
Fix Hub On Device Selection Across Refresh and Download Changes by @Sneakr in #9922
Patch Accelerator.distributed_type as a property, not a bare function by @vineethsaivs in #10028
Stop charging a chat the whole cache because Max Tokens says "Max" by @danielhanchen in #10070
Stop two installer tests answering about the runner instead of the code by @danielhanchen in #10079
Studio: split parallel tool-call arguments at JSON object boundaries by @danielhanchen in #10059
Move the studiobench attribution config in with the harness by @danielhanchen in #10100
Fix Accelerator.distributed_type bound-method bug on single-GPU patch by @mayuriphad in #10072
Wrap Accelerator.distributed_type patch in property() (Closes #10016) by @mihiii1337 in #10087
Add configurable LAN access port by @wasimysaid in #10082
Add GGUF fit-target control and wire to llama-server --fit-target by @aiSynergy37 in #4882
Use the public fastmcp call_tool so the MCP guard test survives fastmcp 4 by @danielhanchen in #10110
Scope the llama.cpp update test's Popen patch to the installer by @danielhanchen in #10105
Run the llama.cpp keep-install back-compat tests on Windows by @danielhanchen in #10106
Stop the torchao ROCm test from depending on the host's GPU by @danielhanchen in #10099
fix(studio): show the confirmation card the argument text the tool is run with by @Lyxot in #10023
Reduce the comments added by PR #9979 by @danielhanchen in #10107
Give the unsloth_zoo pytest step a per-test timeout by @danielhanchen in #10108
Studio: report a context size for models loaded through MLX by @NilayYadav in #10098
Studio: tell the browser about the bootstrap shutdown deadline by @danielhanchen in #10104
AMD: Fix BF16 detection on gfx10 ROCm GPUs by @sunjingwen21 in #7682
fix(amd): gate training GPU selection on the installed torch build's arch coverage by @LeoBorcherding in #8791
Linux: fix ROCm detection on split Debian stacks by @TaskPuppyNatani in #8886
Studio: send tool results to the model in the right order by @NilayYadav in #10091
studio: keep MTP on for GLM-5.3-Flash by @danielhanchen in #10119
Studio: report the GPUs the OS sees, not just the ones PyTorch opened by @danielhanchen in #9858
Apply the context length and quant settings when attaching to a running server by @NilayYadav in #10093
Reuse the MLX prompt cache on hybrid recurrent models by @Lyxot in #10062
Report a trainer kwarg that neither the trainer nor the config takes by @vineethsaivs in #9944
Studio: keep a Downloads entry when the list is empty by @Souravrajvi0 in #9849
Kaggle T4: Latest_compile and Vision_FLA_compile legs, plus gpt-oss completions and GGUF by @danielhanchen in #9700
Kaggle CI: use both accounts, weighted by their own weekly quota by @danielhanchen in #10121
Kaggle T4: split the load phase into fetch and weight load by @danielhanchen in #9692
fix(amd): don't enable torch.compile when Triton's C toolchain can't compile it (Windows) by @LeoBorcherding in #7704
fix(AMD): route Fedora/Bazzite to per-arch ROCm wheels when no version is readable by @LeoBorcherding in #9152
Studio: use plain arrow icons for up, down, left and right by @shimmyshimmer in #10129
Prune an open PR's caches once nothing has read them by @danielhanchen in #10123
Cache the same Playwright engines under the same key by @danielhanchen in #10122
Make the notebooks smoke job reach the check it exists for by @danielhanchen in #10125
Remove OpenSSF Scorecard workflow by @danielhanchen in #10133
feat(studio): give MLX models a real context length, from the model config to the KV cache by @Lyxot in #8125
Bump the actions group with 3 updates by @dependabot[bot] in #9797
Bump the cargo-tauri group across 1 directory with 6 updates by @dependabot[bot] in #9798
Stop unsloth start claude from opening with a model it cannot use by @NilayYadav in #10089
studio: fall back to the MTP/ folder when a repo publishes no root drafter by @danielhanchen in #10124
Studio: auto-generate an admin password for headless public launches by @danielhanchen in #7392
Studio: keep one MCP connection per chat instead of reconnecting every call by @NilayYadav in #10095
Studio: restore Deep Research plan approval by @wasimysaid in #10134
Studio: line up On Device rows, and list partial downloads by @shimmyshimmer in #10128
Widen model picker panel and header controls by @shimmyshimmer in #10132
Studio: keep download restart toasts job-scoped by @wasimysaid in #10136
MLX: drain GPU work before clearing the cache by @Lyxot in #9921
Fix stale Qwen sampling defaults after Studio upgrades by @Imagineer99 in #9643
fix(studio): enforce Hub token caller boundary by @Imagineer99 in #10076
Studio: keep tools available when a picture is attached by @NilayYadav in #10092
Fix pre-commit formatting drift after #10076 by @Imagineer99 in #10154
desktop: check for updates periodically by @mahiatlinux in #9505
studio: use generation usage for conversation search budgets by @mahiatlinux in #9438
studio: fetch the XPU triton wheel with --dest, not pip's -d by @Sletch in #10073
fix(studio): poll inference status while CLI model is loading by @jimdawdy-hub in #5901
Studio: restore sticky checkpoint boundaries on tool-heavy threads by @alkinun in #9971
Studio: surface every live provider model in the picker by @danielhanchen in #5684
Fix llama.cpp Studio update failures on GitHub API rate limits (#9970) by @Souravrajvi0 in #9980
Studio: dispatch on OpenAI agentic web_search action variants by @danielhanchen in #5787
studio: tighten MTP reload guards and asymmetric spec flags for #5582 by @danielhanchen in #5696
Studio: don't re-prompt after model produced a complete answer by @danielhanchen in #5714
Studio: don't size attached documents by a local model on hosted chats by @NilayYadav in #10164
Studio: keep tool cards in place when you edit a reply by @NilayYadav in #10162
Studio: keep a reply's details when you edit its text by @NilayYadav in #10161
Studio: keep both branches when a chat's first message is edited by @NilayYadav in #10163
Stop the kwarg-spacing post-pass dying on Python 3.11 by @danielhanchen in #10137
Studio: stop Deep Research from throwing away a finished report by @NilayYadav in #10166
Studio: show what changed in llama.cpp updates by @oobabooga in #10117
Studio: test that a chat message keeps its identity (#9984) by @umran666 in #10064
Reflow comments in studio/backend/core/inference to 120 columns by @danielhanchen in #10115
Trim comments in studio/backend core services and the unsloth package by @danielhanchen in #10116
Trim comments in the studiobench harness and unsloth_cli by @danielhanchen in #10118
Studio: fix OAuth sign-in for MCP servers like Notion by @NilayYadav in #10167
Studio: align Run Settings and titlebar controls by @alkinun in #10175
Trim comments in studio/backend routes, utils and installers by @danielhanchen in #10112
Update CODEOWNERS for current maintainers and code areas by @danielhanchen in #10191
Security audit: re-approve fastmcp-slim 4.0.x and unsloth-zoo 2026.8.17 by @danielhanchen in #10187
Stop the research cancellation tests failing on a busy CI runner by @danielhanchen in #10190
Fix the frontend unit tests broken by a merge skew between #10161 and #10162 by @danielhanchen in #10186
Fix the two stale source-text contract tests blocking every PR by @danielhanchen in #10185
Fix Studio preferring MTP sidecars over embedded heads by @wasimysaid in #10181

New Contributors

@milewski made their first contribution in #9539
@DhruvaMyakeri made their first contribution in #9913
@Greninja44 made their first contribution in #9911
@mayuriphad made their first contribution in #10072
@mihiii1337 made their first contribution in #10087
@aiSynergy37 made their first contribution in #4882
@sunjingwen21 made their first contribution in #7682
@TaskPuppyNatani made their first contribution in #8886
@Sletch made their first contribution in #10073

Full Changelog: v0.1.804-beta...v0.1.806-beta

unslothai/unsloth v0.1.806-beta 2x Faster Qwen3.8-Flash + GLM-5.3-Flash MTP on GitHub