unslothai/unsloth v0.1.801-beta on GitHub

Thanks for the support for Qwen3.8-27B and Unsloth Desktop last week! For this release, we merged 200+ PRs to introduce many new features, fixes including:

Auto Compaction (Experimental) for longer chats beyond context limits
Remote & LAN Access (Preview) for easy network access without Cloudflare links
Faster Chat - Improved streaming performance, reduced UI lag, and smoother long conversations.
Support for custom llama.cpp builds. Toggles for Cache RAM, Mmap, Mlock, Checkpoints, Spe Decoding KV Cache, Vision On / Off
Unsloth Dynamic v3.0 is released. New Qwen3.8-27B Dynamic v3.0 GGUFs deliver >10% higher top-1 accuracy compared to everyone else. Works with Unsloth.

Auto compaction (Experimental)

You can now do long chats and surpass a model's context limit, with evicted turns
remaining searchable.

Long local chats can continue past the context limit by rolling older turns out of
active context.
Auto compaction creates fresh context epochs instead of permanently trimming replies.
Evicted conversations remain searchable, with improved fact recall.
Context window information is shown before chat begins.

How it works

Entire oldest turns are removed only when needed, never mid-message. The saved transcript stays unchanged.
Evicted turns are indexed into a per-thread searchable archive using Unsloth's existing RAG pipeline (store, chunker, embedder, retrieval). Lexical search is prioritized because chat recall is usually exact matches like names, numbers, or IDs.
A recall is forced during eviction instead of relying on the model to search itself. Later retrieval uses search_conversation.
The archive persists across epochs, allowing future compactions to recover previously evicted context.
No summarization: it showed little benefit and added ~190s per compaction.

Remote & LAN access (Preview)

Access Unsloth from another device on your network, managed from settings.

New dedicated Settings section for remote access.
Enable or disable LAN access without restarting.
Supports connection addresses, QR codes, and optional auto-start.
LAN access is disabled by default and requires changing the generated admin password.

Chat + Functionality improvements

Long chats are faster with optimized streaming and thread handling.
Projects organize chats, files, and workspaces together.
Chats remember composer settings, system prompts, and model sampling options.
Added prompt queueing and customizable keyboard shortcuts.
Added edit_file tool for partial file modifications.
Improved tool calling and MCP behavior.

Hardware + inference improvements

Support for custom llama.cpp builds.
Toggles for advanced settings - Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, Vision On / Off
Better validation for missing GPU support and oversized GGUF models.
Improved MTP performance and VRAM handling.
Better compatibility across ROCm, xFormers, and flash-attention setups.
Added Intel XPU support.

API + serving

Responses API supports structured output formatting.
Improved recovery when llama-server crashes.
Added OpenCode V2 support for unsloth start.

Hub + Desktop + training

Hub supports more image/video workflows.
Faster and improved GGUF downloads.
Desktop packaging and permissions improved.
Training scheduler and preprocessing improvements.

Platform	Link
Windows	Download
macOS	Download
Linux / Ubuntu (deb)	Download
Linux (AppImage)	Download
Linux (Arm64)	Download

What's Changed

Bump install.sh / install.ps1 pin to unsloth>=2026.8.18 by @danielhanchen in #8836
Give the llama-flag fixtures the parallelSlotsClamped they now require by @vineethsaivs in #8837
Studio: preserve pre-tool reasoning inside the GGUF tool loop by @losuler in #8581
docs: fix dead Linux .deb download link by @yzxcj797 in #8891
Model hub: show the Meta mark on the Unsloth re-upload of Muse Glimmer by @shimmyshimmer in #8949
Studio: speed up the local model inventory and keep it off the API loop by @oobabooga in #8770
Studio: reuse the cached GGUF work a load request has already done by @oobabooga in #8771
studio: keep each tool-call round separate when the delta index restarts by @mahiatlinux in #8754
studio: route bare tool-call fragments to the call that owns the index now by @mahiatlinux in #8755
studio: opt-in model auto-switch for the image and video apis by @mahiatlinux in #8766
Studio: keep project chats in Recents, add a Projects sidebar section by @shimmyshimmer in #8932
Studio: find the real Documents folder on Windows for project workspaces by @shimmyshimmer in #8955
fix(unsloth studio): only preprocess the rows a max_steps run will actually use instead of preprocessing validated datasets by @LeoBorcherding in #8890
Studio: stop the chat UI falling behind a fast stream by @oobabooga in #8845
studio: keep each chat's composer pills and settings with the chat by @mahiatlinux in #8686
studio: allow a per-connection max tokens limit on every provider by @mahiatlinux in #8944
Crash guard: read dumpability in order, and stop losing crashes to scope and naming by @danielhanchen in #8808
Studio: size the max_steps row bound for the data-parallel world by @danielhanchen in #8958
Studio: make the paste-to-attachment threshold a chat setting by @shimmyshimmer in #8963
Fix CI on main: stale test doubles, a stale router stub, and two source defects by @danielhanchen in #8956
Make the startup profile a gate, with budgets from its own measurements by @danielhanchen in #8965
Keep xFormers working when flash-attn 4 is installed, and guard the varlen int32 overflow by @danielhanchen in #8957
Stop six backend tests waiting on the clock instead of on a signal by @danielhanchen in #8975
CI: stop three workflows paying for work they throw away by @danielhanchen in #8976
Give the pasted-text import an extension so the node test runner can load it by @danielhanchen in #8981
Make the packed-boundary guard reachable on the fused cross-entropy path by @danielhanchen in #8959
Studio: Fix embedded MTP performance under partial GPU offload by @oobabooga in #8875
Studio: tokenize the dataset online for plain-text single-pass runs by @danielhanchen in #8960
Studio: keep K and V equal when the flash-attention retry resets the V cache by @danielhanchen in #8827
Studio: hold the Auto-layers context exemption only while --fit runs by @danielhanchen in #8825
Studio: stop a sent prompt staying in the chat composer by @shimmyshimmer in #8849
Chat: queue prompts with Cmd/Ctrl+Enter and drag to reorder the queue by @shimmyshimmer in #8952
Studio: repin the permission-mode contract to behaviour, not a declaration by @danielhanchen in #8986
Cut Repo tests (CPU) roughly in half: stop the synthetic vLLM tests reaping a fake server by @danielhanchen in #8993
Core: sweep every transformers model_type across a process pool by @danielhanchen in #8995
Repair the Tauri retry test against the download-marker pipeline by @danielhanchen in #8996
Studio: read every launcher world size when resolving a step-capped run's passes by @danielhanchen in #9000
Repair the prompt-queue contract test against the queueing refactor by @danielhanchen in #9026
studio: stop prompting to stop chats for a model already loaded by @mahiatlinux in #8943
Studio: remember chat parameters per model by @shimmyshimmer in #8757
Studio: share attached files across a project by @shimmyshimmer in #8756
Fix a 500 on dictation Unload before any backend is resident by @danielhanchen in #9013
Run the repo CPU test suite on all four runner cores by @danielhanchen in #9019
Scan package archives across cores instead of one at a time by @danielhanchen in #9024
Stop TestProbeGating asserting a tier the mapping check is meant to upgrade by @danielhanchen in #9029
Stub the implicit STT load in the disconnect-cancellation test by @danielhanchen in #9031
Stop test_begin_load_never_refuses_auto racing the thread it is asserting about by @danielhanchen in #9034
Measure the streaming stripper against CPU time, not wall clock by @danielhanchen in #9035
Fix silent lr_warmup_steps no-op under the default constant scheduler by @Eoin-Houstoun in #8464
Repair the contract tests #8943 left behind on main by @danielhanchen in #9067
Take the loaded-models indicator suite off the UI job's critical path by @danielhanchen in #9060
Keep the killed-by-signal exit status when release() fails by @danielhanchen in #9072
Put the event-loop latency bound between the two things it separates by @danielhanchen in #9077
Stub unsloth in test_safetensors_reasoning_stream so it can run first by @danielhanchen in #9027
Report what the launcher logged when a signal test fails by @danielhanchen in #9079
Studio: let Allow microphone recover from a saved "Don't allow" on Windows by @shimmyshimmer in #9006
Run both ends of the interpreter matrix on a pull request, all four on main by @danielhanchen in #9080
fix(studio): enable microphone/camera access in Linux WebKitGTK webview by @rsd-darshan in #8720
Studio: budget the JavaScript that runs before the first screen by @danielhanchen in #8964
fix(studio): derive disk percentage from displayed usage by @dyuan311 in #9021
Studio: stop rereading the whole reply on every streamed arrival by @danielhanchen in #9012
Studio: stop every message part re-rendering on each streaming chunk by @danielhanchen in #9014
Studio: stop a LaTeX rewrite throwing away the whole retained markdown prefix by @danielhanchen in #9017
Studio: stop later code spans changing whether an earlier one is escaped by @danielhanchen in #9028
Chat: stop a message delete from re-rendering the whole thread by @danielhanchen in #9042
Chat: stop a keystroke in the composer costing one pass over the thread by @danielhanchen in #9054
Let every main run finish instead of cancelling it before it starts by @danielhanchen in #9082
One multi-turn smoke script for all three operating systems by @danielhanchen in #9086
Studio: give a CRLF reply the same incremental rendering as an LF one by @danielhanchen in #9088
Studio: sweep the streaming markdown corpus at every prefix by @danielhanchen in #9085
Do not let the signal handler's own logging stop it dying by @danielhanchen in #9083
Stop the media switch test racing a 0.3s budget against 2s of cold path by @danielhanchen in #9097
Studio: honor the Responses API text.format structured output by @NilayYadav in #8846
Studio: stop the streaming prefix checks scanning the whole reply by @danielhanchen in #9038
Studio: stop the streamed reply being flattened on every arrival by @danielhanchen in #9049
Studio: strip the trailing template placeholder once, on the finished reply by @danielhanchen in #9101
Studio: fix the Windows-only path failures in the frontend test suite, and catch the shape from Linux by @danielhanchen in #8980
Studio: run the frontend unit tests on Windows in CI by @danielhanchen in #9099
Run the Backend CI matrix in parallel, like its sibling job already does by @danielhanchen in #9095
Pair the diffusion warmup presets with a scheduler that uses them by @vineethsaivs in #8593
Make the models a defaults file claims actually load it by @vineethsaivs in #8726
Derive warmup steps from a model default that gives a ratio by @vineethsaivs in #8723
Studio: let the hub run non-GGUF image and video models by @NilayYadav in #8855
Restore crypto.randomUUID before the Studio bundle evaluates by @Lyxot in #9075
studio: refuse a gguf that cannot fit in free vram plus available ram by @mahiatlinux in #8883
studio: add settings-managed LAN access by @mahiatlinux in #8951
Ask the repo, not sys.path, whether a stubbed name is third-party by @danielhanchen in #9133
Stop the sidecar tests writing into the real Studio install by @danielhanchen in #9131
Delete a test that never ran, and stop another appearing by @danielhanchen in #9135
Make a hung kaggle launcher say where it is stuck by @danielhanchen in #9104
Stop the banner layout suite waiting out the update timer on every boot by @danielhanchen in #9105
Give each xdist worker its own torch.compile cache directory by @danielhanchen in #9134
One interpreter leg on a pull request, and a floor lint that reads more than syntax by @danielhanchen in #9100
Split Chat UI Tests into four shards on the Studio boundaries by @danielhanchen in #9132
Bundle Unsloth GGUF CI onto one runner, matching Windows and macOS by @danielhanchen in #9139
Give the small Windows checks one box per image instead of one box per check by @danielhanchen in #9143
Let tests skip the settle wait between retried VRAM reads by @danielhanchen in #9141
Studio: keep the streaming render harness the perf PRs kept rebuilding by @danielhanchen in #8969
Studio: pin why a Streamdown remount keeps its highlighted code by @danielhanchen in #9048
Chat: cover which messages a delete takes with it by @danielhanchen in #9087
Studio: pin what the trailing-placeholder window does with nested placeholders by @danielhanchen in #9091
Studio: stop a settings module reading a chat store key before it exists by @danielhanchen in #8979
Studio: load the settings tab panels when they are shown, not at launch by @danielhanchen in #8966
Name the encoding when reading wslinfo's output by @danielhanchen in #9142
Studio: keep pandas out of the backend startup import graph by @danielhanchen in #8962
Run the three loaded-models-indicator engines at once by @danielhanchen in #9158
Unblock Mac Studio GGUF CI, red on every main run since #8883 by @danielhanchen in #9155
Drop the Exa preset from the MCP menu by @shimmyshimmer in #9160
Studio: stop the overlay rail going click-through while it is scrolling by @danielhanchen in #9150
Open the connection form when Settings > Connections has nothing to list by @shimmyshimmer in #9157
Security audit: pin openai, and re-review the four digest-pinned entries by @danielhanchen in #9148
Say why a Xet download sits at 0% until it finishes by @shimmyshimmer in #9159
Stop running every macOS workflow on every commit to main by @danielhanchen in #9174
Time the install from CI, without changing the installers by @danielhanchen in #9153
Studio: measure where a heavy thread stalls, across engines and thread size by @danielhanchen in #9016
Studio CI: make a failing browser smoke say why, and stop it skipping the build gates by @danielhanchen in #8983
Stop spending the Actions cache budget on caches nothing reads by @danielhanchen in #9151
Fix low VRAM mmproj fallback in Studio by @wasimysaid in #9173
Run two short jobs as background lanes of Lint CI, not on their own runners by @danielhanchen in #9176
Studio: fail fast when the installed llama.cpp prebuilt has no kernels for this GPU by @NilayYadav in #8841
Studio: bulk chat management multi-select move/pin/archive/export/delete by @NilayYadav in #8745
hide app-only MCP tools from the model and explain why local commands are disabled by @NilayYadav in #8739
Studio: remember last used model in the backend, not per-browser by @NilayYadav in #8587
Studio: show MCP server display names instead of internal ids by @NilayYadav in #8557
Studio: make the chat thread stop getting slower as it fills by @danielhanchen in #8992
Studio: recover marked reports across reasoning channels by @alkinun in #9119
show the local model's context window before the first token count by @mahiatlinux in #8882
studio: stop nudging a turn that asks the user for a missing detail by @mahiatlinux in #8946
Say both load fallbacks when both fire, and repair the two suites that guard it by @danielhanchen in #9189
Run the Mac GGUF inference phases in the Mac UI job, not on a second runner by @danielhanchen in #9213
Print what the server said when an inference smoke request 4xx's by @danielhanchen in #9202
Allowlist huggingface-hub's http_backoff loop so Security audit stops failing by @danielhanchen in #9252
Count a capped overlay stack by what it reads, not by how the cap is spelled by @danielhanchen in #9253
Studio: support llama.cpp semantic version output by @oobabooga in #9127
Bound every apt step in CI, and fix the retry that never ran by @danielhanchen in #9256
Make apt fail fast, so the retry has something to retry by @danielhanchen in #9260
Record the venv own requirement digests, not the installers by @danielhanchen in #9263
Stub all three fields the snapshot reads, not two of them by @danielhanchen in #9274
Ask whether Playwright system libraries are missing before installing them by @danielhanchen in #9273
Stop blaming the recipe for a turn that never came back by @danielhanchen in #9266
Move the two version-incidental macOS jobs onto the emptier pool by @danielhanchen in #9277
Give the Colab freeze a pip cache, like the job beside it already has by @danielhanchen in #9276
Wait for the Compare nav instead of sampling for it by @danielhanchen in #9264
Install the browser engines each Chat UI shard actually drives by @danielhanchen in #9283
Update README to include Intel XPU support description for unsloth studio by @JoshuaL3000 in #9250
Allow unlimited Deep Research time budgets by @alkinun in #9078
Stop the orphan-scan tests asking the host whether an invented PID is alive by @danielhanchen in #9280
Repair the tests main is currently red on by @danielhanchen in #9192
Re-type the native default, not whatever the input happens to show by @danielhanchen in #9282
Studio: confirm an Audio model swap instead of dead-ending on the backend 409 by @shimmyshimmer in #9092
fix(studio): refuse a hand-set context unified memory cannot hold by @danielhanchen in #9172
Studio: size Xet download buffers from free RAM, not total by @shimmyshimmer in #9052
Studio: make the partial download affordance tell the truth by @shimmyshimmer in #8982
Cache uv's downloads, which is now the largest single cost in CI by @danielhanchen in #9271
Studio: incrementally tokenize streaming code fences by @oobabooga in #8935
Desktop: ship a complete Linux AppImage by @oobabooga in #9113
Studio: link the Read aloud TTS row to Audio and show the preview generate wait by @shimmyshimmer in #9093
studio: show the model a chat was started on, and offer it back by @shimmyshimmer in #9081
Desktop: make every drop zone take a drop again (#9036) by @shimmyshimmer in #9056
Studio: say when a scan folder cannot be read instead of showing no models by @shimmyshimmer in #9053
Studio: recover the OpenAI passthrough from a crashed llama-server by @shimmyshimmer in #8999
Studio: stop the desktop health watchdog killing a backend that is busy generating by @shimmyshimmer in #8990
Studio: add a keyboard shortcuts page and make the existing chords rebindable by @shimmyshimmer in #8948
Scroll into view before a forced click by @danielhanchen in #9291
studio: keep each chat's sampling params and system prompt with the chat by @shimmyshimmer in #9055
Run the zoo suite in parallel, minus the two files that cannot share a worker by @danielhanchen in #9285
Cache the .deb set webkit needs, so a dead mirror costs a miss not a shard by @danielhanchen in #9289
Revert "Say why a Xet download sits at 0% until it finishes" by @shimmyshimmer in #9293
Default preserve thinking on for Qwen3.8 by @Imagineer99 in #9096
Studio: price a partial GGUF by what is left to fetch by @shimmyshimmer in #8989
Studio: add an edit_file tool so agents stop rewriting whole files by @shimmyshimmer in #8753
fix(studio): img2img/inpaint on GGUF image models — from_pipe fallback with no recast (#9186) by @yzxcj797 in #9193
Install Studio on a Windows host whose C# compiler cannot run by @danielhanchen in #9178
fix(AMD): retry ROCm llama-server with bundled HIP on HIP/ROCR mismatch by @indrajeetapache in #9002
Studio: use a custom llama.cpp build by @Etherll in #9292
Studio: keep health checks responsive during Anthropic generation by @Imagineer99 in #9201
Studio: propagate required backend version to repair pipeline (#8610) by @karan68 in #8670
Keep a sampler's std and uuid format across a save and reopen by @vineethsaivs in #9225
Studio: never pick a macOS AppleDouble sidecar as a GGUF by @sts-change in #8919
Studio: follow up #9201 across non-streaming sibling paths by @oobabooga in #9312
Studio: take common SQLite reads off the event loop thread by @oobabooga in #9234
Keep xFormers attention masks on the GPU running each layer by @oobabooga in #8516
Studio: add rolling context windows for local GGUF chat by @alkinun in #8961
Studio: keep and search the turns rolling context evicts by @danielhanchen in #9074
Studio: recall the latest version of a fact, not the most quotable one by @danielhanchen in #9161
Studio: compact a chat by resetting the epoch, not by trimming it forever by @danielhanchen in #9162
Support OpenCode V2 in unsloth start by @pascalandr in #9275
Keep path_utils off PEP 604, which the 3.9 floor gate rejects by @danielhanchen in #9335
Stub the two helpers the sliced harnesses started importing by @danielhanchen in #9342
Put back the AppleDouble GGUF filters #9074 reverted, and guard them by @danielhanchen in #9353
Fix the two backend guards that #9292 turned red on the 3.13 leg by @danielhanchen in #9348
Drain the sampling debounce long enough for the node CI actually runs by @danielhanchen in #9332
Take two ubuntu jobs off their own runners: absorb one, delete the other by @danielhanchen in #9360
Studio: stop docked composer from squeezing download panel by @oobabooga in #9334
Drain the sampling suites on the store's write chains, not on node's loader by @danielhanchen in #9367
Say which runtime components an AppImage is missing before its env hygiene by @danielhanchen in #9370
Drop the HunyuanImage GGUF the Hub no longer serves by @danielhanchen in #9369
Studio: fix two intermittently failing tests by @danielhanchen in #9200
Acknowledge the Colab rotation the oracle snapshot is measuring by @danielhanchen in #9376
Studio: carry a restated instruction once, not eight times by @danielhanchen in #9368
Stop 13 jobs rebuilding the same frontend on every commit by @danielhanchen in #9375
Stop 19 test files racing on one PowerShell startup cache by @danielhanchen in #9371
Stop the pip cache key duplicating itself on every Python patch bump by @danielhanchen in #9377
Studio: list Ollama models in the chat picker and load their manifest refs by @lonexreb in #9237
Stop the Studio temp probe leaving anything behind by @danielhanchen in #9298
Stop the dist cache key hashing 457 files the rebuild check never reads by @danielhanchen in #9380
Studio: overturn a chat-only MLX verdict the stack contradicts (#9120) by @yzxcj797 in #9124
Put the managed Node bin dir on PATH when spawning stdio MCP servers by @NilayYadav in #9304
Studio: normalize replayed tool-call ids for external providers by @NilayYadav in #9116
Don't select Flash Attention 2 for float32 model loads by @NilayYadav in #9110
Studio: keep every cached revision's quants on the on-device card by @Lyxot in #9381
Studio: put Remote and LAN access on their own settings tab by @shimmyshimmer in #9389
Run six pinned-symbol jobs on one runner instead of six by @danielhanchen in #9385
Studio: take the model name back off the chat rows by @shimmyshimmer in #9395
Stop five Windows jobs rebuilding the same frontend on every commit by @danielhanchen in #9387
Run the Windows UI suites in two parallel lanes instead of five in sequence by @danielhanchen in #9391
Studio: keep a downloaded model listed when upstream moves on by @oobabooga in #9347
Studio: offer only the Codex models the ChatGPT plan can actually reach by @NilayYadav in #9302
Studio: place the vision projector on the CPU only when it does not fit in VRAM by @danielhanchen in #9063
Fix Qwen3.8 presence penalty defaults by @Imagineer99 in #9372
Studio: Forward thinking controls and return reasoning on /v1/messages and /v1/responses by @NilayYadav in #9107
Studio: keep the working llama.cpp install when a prebuilt update can't move it aside by @NilayYadav in #9221
Studio: add a Disable Vision toggle to a model's Advanced Settings by @danielhanchen in #9383
Studio: stop the scroll stabilizer invalidating style for the whole thread by @danielhanchen in #9328
Studio: never offer a calibration imatrix as a GGUF variant by @danielhanchen in #9409
Stop Codex cancelling its own first turn while llama-server reads the prompt by @danielhanchen in #9403
Studio: stop a reset dropping the task when the user types short prompts by @danielhanchen in #9379
Studio: finish a reply that hit Max Tokens instead of asking whether to by @danielhanchen in #9382
Format the mmproj placement policy tests so pre-commit stops failing every PR by @danielhanchen in #9414
Studio: scope the sidebar drag's custom-property writes off the thread's ancestors by @danielhanchen in #9400
Studio: size a large tool result to the window the model is running with by @danielhanchen in #9384
Report what the model server actually said when a stream fails by @danielhanchen in #9390
Admit generations against the KV cache, not just the slot count by @danielhanchen in #9392
Studio: add Mmap/Mlock, draft KV cache dtype, Checkpoints and Cache RAM to Run settings by @danielhanchen in #9410

New Contributors

@losuler made their first contribution in #8581
@yzxcj797 made their first contribution in #8891
@Eoin-Houstoun made their first contribution in #8464
@dyuan311 made their first contribution in #9021
@indrajeetapache made their first contribution in #9002
@karan68 made their first contribution in #8670
@sts-change made their first contribution in #8919
@pascalandr made their first contribution in #9275
@lonexreb made their first contribution in #9237

Full Changelog: v0.1.800-beta...v0.1.801-beta

unslothai/unsloth v0.1.801-beta Auto compaction (preview) + LAN Remote Access on GitHub