unslothai/unsloth v0.1.802-beta on GitHub

Thanks for the support for Qwen3.8-27B and Unsloth Desktop! This is a bug fix release with 170+ PRs.

MLX fixed - Some MLX and Mac runtimes did not run correctly
LAN API keyless / password-less is now supported
XET / HTTP download toggle - clearer download progress
AMD bug fixes for Strix Halo, all RDNA GPUs + 170 bug fixes

Features

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.

What's Changed

Bump install.sh / install.ps1 pin to unsloth>=2026.8.19 by @danielhanchen in #9422
Say what killed a UI wrapper script instead of reporting a bare 143 by @danielhanchen in #9407
Fix stale libappindicator3-dev guard blocking desktop releases by @danielhanchen in #9426
studio: drag and drop files onto project sources and knowledge bases by @mahiatlinux in #9147
Read a llama install's ggml tree from its release when the marker lacks one by @danielhanchen in #7912
studio: add custom stt connections by @mahiatlinux in #9349
Repair the two studio test suites #9410 left behind by @danielhanchen in #9437
CI: hold the Anthropic SDK below 1.0 in the inference smoke workflows by @danielhanchen in #9432
studio: collapse the reasoning pane with grid rows instead of a measured height, behind a flag by @danielhanchen in #9423
Studio: turn on the sidebar drag's scoped custom-property writes by @danielhanchen in #9441
Studio: re-ask the projector placement after a tensor-parallel downgrade by @danielhanchen in #9408
Studio: hide the Tensor Parallelism switch on diffusion models by @danielhanchen in #9418
Exclude the imatrix from the local models route too by @danielhanchen in #9411
Studio: fall back to the legacy --swa-checkpoints spelling by @danielhanchen in #9416
Retry the desktop ownership id read that a concurrent publish denies by @danielhanchen in #9459
Bound the SDK majors CI installs, and guard the ones the probes call by @danielhanchen in #9460
Stop the chart metric formatters truncating values of 1000 and up by @vineethsaivs in #8724
Studio: queue media generation behind model teardown by @divagr18 in #8866
Studio: keep server chats visible when legacy IndexedDB stalls by @oobabooga in #9446
fix(studio): rewind the MLX PRNG in the KV quantization probe without item assignment by @Lyxot in #9478
install: fetch managed Python when distro uv.toml sets downloads=manual by @gokay-ai in #9358
Studio: admit GGUFs that fit carved-out Vulkan UMA by @alkinun in #9461
Studio: settings nav badges and the Shortcuts icon by @shimmyshimmer in #9497
studio: stop a menu dismissal deleting a chat message by @mahiatlinux in #9243
Unwedge Backend CI: one deadlocked test, and a timer that names the next one by @danielhanchen in #9515
Studio: copy a thread selection without building the styled clipboard flavour by @danielhanchen in #9457
Studio: turn on the grid-based reasoning collapse by @danielhanchen in #9511
Studio: defer off-screen code fence highlighting, one way only by @danielhanchen in #9462
studio: measure the code block collapse the thread override prevents by @danielhanchen in #9464
Studio: accept .ods files from drag and drop by @Etherll in #9455
open every attachment type in a preview modal by @mahiatlinux in #8655
Import the OpenDocument MIME constants from where they are exported by @danielhanchen in #9520
Studio: don't let MCP training kill a live chat stream by @NilayYadav in #9434
Desktop: attach dropped source files on desktop instead of refusing them by @NilayYadav in #9495
Studio: avoid AppImage COLRv1 font crashes by @wasimysaid in #9473
Studio: stop an unterminated code fence re-repairing itself on every chunk by @danielhanchen in #9517
Studio: list the other embedding models, and show Documents & RAG in Data by @shimmyshimmer in #9514
Studio: document reasoning on the Agents tab, and add --reasoning-effort by @shimmyshimmer in #9513
Bound the backend pytest session so a wedged xdist worker cannot eat the job by @danielhanchen in #9530
Studio: paint a long thread's newest messages immediately instead of rebuilding all of them first by @danielhanchen in #9058
Allowlist the remote-server locale keys in the desktop branding sweep by @danielhanchen in #9529
Give the Studio T4 leg a Kaggle client that can read its own credential by @danielhanchen in #9535
Assert the event-loop property directly in the two load_freeze latency tests by @danielhanchen in #9532
studio: stream event feeds over post by @mahiatlinux in #9533
Studio: stop the KV starvation message reading as a context limit on the client by @danielhanchen in #9417
Studio: measure a dense tool result instead of estimating it, and never raise a configured cap by @danielhanchen in #9421
Studio: stop a failed markdown chunk from taking the whole app down by @danielhanchen in #9236
Write down the rule for UI changes to Studio, and put it on every pull request by @danielhanchen in #9463
Skip llama-server discovery when there is no launch revision, and scope the VRAM finalizer assertion by @oobabooga in #9325
Desktop: stop sending OS notifications for model loads by @oobabooga in #9494
Say "Unsloth" or "Unsloth Studio" instead of bare "Studio" in settings by @shimmyshimmer in #9526
Revert "Write down the rule for UI changes to Studio, and put it on every pull request" by @danielhanchen in #9544
Studio: re-run admission when a cancelled waiter was blocking the line by @danielhanchen in #9415
Installer: replace an existing venv whose interpreter is gone by @oobabooga in #9501
Studio: repair duplicate package metadata during updates by @oobabooga in #8515
Never carry the opening request without the turn that follows it by @danielhanchen in #9424
studiobench: a real-path performance benchmark and A/B simulator for Studio by @danielhanchen in #9296
Studio: stop two tabs both auto-continuing the same truncated reply by @danielhanchen in #9425
Prevent Studio reloads from flashing blank by @oobabooga in #9251
Studio: escape the dependency fast path when AMD torch is the wrong wheel by @oobabooga in #9499
Studio: name the GPU agent, not the CPU, in the rocminfo probe by @oobabooga in #9498
fix(export): forward private repo flag to push_to_hub_gguf in CLI and Studio by @Siva010 in #9474
Studio: stop compaction spending the room the compacted turn needs by @oobabooga in #9442
studio: normalize ollama reasoning deltas by @mahiatlinux in #9504
Studio: stop a cancelled prompt from stranding two user turns in chat history by @NilayYadav in #9489
Studio: tray toggle reflects server state by @NilayYadav in #9488
Studio: turn LoRA finetune targets on by default by @NilayYadav in #9436
Studio: keep unrunnable speech GGUFs out of the model pickers by @NilayYadav in #9433
Studio: record media API traffic in the monitor and add verbose_json transcriptions by @NilayYadav in #9217
Studio: a transient training 409 should not poison the start_request_id by @danielhanchen in #9536
Fix delegated generation flag cleanup by @pascalandr in #9469
Studio: re-fit the context when Auto drops serving slots by @oobabooga in #9345
Studio: show web search images inline in chat by @NilayYadav in #9490
Pack the Kaggle T4 legs into one kernel so Studio can use the other session by @danielhanchen in #9538
fix(images): refuse stale-model parameters in /v1/images/generations by @yzxcj797 in #9451
Studio: keep one copy of the embedding model setting by @shimmyshimmer in #9531
studiobench: the verdict tools, so a performance number can be shown to be real by @danielhanchen in #9297
Studio tests: follow the helpers a sliced harness calls, unblocking CI on every open PR by @danielhanchen in #9574
Tests: extract setup.sh's fast path without depending on if vs elif by @danielhanchen in #9579
studiobench: give the action bar selftest's two clocks one origin by @danielhanchen in #9578
Studio: say which part of the prompt does not fit the context window by @danielhanchen in #9413
Studio: stop the client blaming a turn for the tool catalogue beside it by @danielhanchen in #9555
Studio: pay for the tool result framing probe once per model, not once per result by @danielhanchen in #9541
Studio: defer off-screen code fence highlighting by default by @danielhanchen in #9567
Pin what the MLX self-heal installs, and correct the mlx-lm floor by @danielhanchen in #9557
Harden the Studio MLX PRNG rewind against words that are not uint32 by @danielhanchen in #9558
Studio: improve Auto context selection and share remembered model settings by @oobabooga in #9492
Extract the rollback helpers' dependency closure instead of a hand-written list by @danielhanchen in #9592
Studio: paste without formatting always fills the message box by @shimmyshimmer in #9545
Fix main: read the settings route's contract off its parse tree, not its text by @danielhanchen in #9591
Studio: do not call the host CPU-only before the System tab has read anything by @LeoBorcherding in #9315
Fix main: price the serving-slot reduction at the fit floor, not the Auto offload context by @danielhanchen in #9594
Bump the actions group across 1 directory with 6 updates by @dependabot[bot] in #9559
studiobench: put math in the corpus, and refuse to pool across corpora by @danielhanchen in #9341
Studio: size a tool result against the room the thread has left by @danielhanchen in #9563
studiobench: measure what streaming one character costs, and fix the gap windows that mislabel it by @danielhanchen in #9351
Studio: stabilize download speed and ETA (#9388 + #9397) by @danielhanchen in #9593
Studio: give a resize drag its cursor and hit test from an overlay, not the whole document by @danielhanchen in #9598
Read quoted annotations in the import lint, and build the video load doubles off the real backend by @danielhanchen in #9601
studiobench: measure the composer click instead of dying on it, and the 500K and 1M rungs by @danielhanchen in #9356
studiobench: a readiness gate that can admit a windowed thread, and three instruments that were charging their own cost to the app by @danielhanchen in #9439
fix(studio): read the MLX reasoning prefill mode from the rendered generation prompt by @Lyxot in #9288
Studio: keep project chat runs alive across view switches by @NilayYadav in #9129
Fix silent dropping of embed_tokens/lm_head when listed in target_modules by @chakshu-dhannawat in #9331
Restore symlink members when extracting the sd.cpp prebuilt by @FrancescoBrunoDev in #9270
fix(hardware): report driver-level free VRAM in the GPU summary by @yzxcj797 in #9362
add auto-compaction and lan access to readme by @mahiatlinux in #9613
Give huggingface_hub 1.18+ back its resumable HTTP partials by @danielhanchen in #9602
Contain the installer tests' writes to the shared venv root (fixes the AMD fastpath probe flake) by @danielhanchen in #9587
studio: stop the sandbox workdir aliasing /tmp in git bash by @mahiatlinux in #8897
studiobench: settle the instruments that published four wrong numbers by @danielhanchen in #9542
desktop: disable the webkitgtk dmabuf renderer on nvidia, not only on wayland by @mahiatlinux in #8884
Fix fbgemm blockwise FP8 crashes on unsupported shapes and blocksizes by @hellopahe in #8953
Stop the corner update toasts being cut off at the bottom by @shimmyshimmer in #9246
Stop gating CI on README prose in the RDNA 1 guards by @danielhanchen in #9619
Studio: honor the OpenAI sampling parameters on the MLX backend by @Lyxot in #9262
Survive transformers 5.x configs that refuse a global per-layer read by @danielhanchen in #8492
studio: align tool-call nudge defaults across loops by @Biotrioo in #9126
fix(studio): auto-enable --embedding for GGUFs missing pooling_type by @Souravrajvi0 in #9149
Studio: support external OpenAI-compatible TTS endpoints for Read Aloud by @NilayYadav in #9214
CI: keep the host-default README guard scoped to Studio docs by @Apoze in #9623
Fix Apertus Instruct model mapper pointing at the base upstream repo by @vineethsaivs in #7339
CI: give the overlay rail harness the vite it needs, and tighten the wildcard-host detector by @danielhanchen in #9654
fix(studio): point denied --parallel extra args at the n_parallel load field by @yzxcj797 in #9512
Give start_vite's unit tests a toolchain, rather than weakening the refusal by @danielhanchen in #9656
Prefer the .cmd sibling when Windows resolves an extensionless npm shim by @lonexreb in #9238
CI: guard the host default structurally, not by README heading by @danielhanchen in #9655
Studio: stop dropping quantized KV cache on tensor split by @kissa01 in #8939
studiobench: a window's stream-cost integrity is about one response by @danielhanchen in #9609
Make the legacy-migration picker step deterministic, and unit-test the migration by @danielhanchen in #9603
Stop to_sharegpt blanking the input column when no merged_prompt is given by @vineethsaivs in #8277
Say why a Xet download sits at 0%, in a toast that clears the hub toolbar by @danielhanchen in #9658
studiobench: two ui parity gates that could only ever say yes by @danielhanchen in #9608
Make the desktop signing secrets gateable, and decode the cert safely by @danielhanchen in #9657
lint: refuse a name bound twice in one scope of a changed Python file by @danielhanchen in #9618
studiobench: bound the scroll gesture at both ends, at the extent's own allowance by @danielhanchen in #9610
studiobench: refuse the message that is still being written, and score the rest of the thread by @danielhanchen in #9575
fix(logs): include the backend session log in the support diagnostics report by @LeoBorcherding in #8877
studiobench: refuse a streamed tail the corpus cannot deliver by @danielhanchen in #9659
Studio: put the notification rail back in its bottom-right corner by @shimmyshimmer in #9547
Studio: drop the duplicate download toast on a hub auto-load by @shimmyshimmer in #9663
Stop the installer rolling back a successful install when a shell profile cannot be written by @Lyxot in #9269
Cut the host-defaults installer windows from structure, not from comments by @danielhanchen in #9662
Studio: keep SSM kernel installs alive and reap the worker on load timeout by @indrajeetapache in #9419
Keep the real quantization skip list on save, and add an opt-in multi-GPU device map by @danielhanchen in #8968
fix(logs): quiet the polls and drop the duplicated access lines by @LeoBorcherding in #8763
Xet notice: one toast carrying both messages, three times per install, gone when the download is by @danielhanchen in #9660
fix chat search dialog stutter on open by @mahiatlinux in #8514
Studio: scale, crop and trim MiniMax-H3 reference media by @oobabooga in #9631
fix(amd): report AMD VRAM on Windows by joining the adapter counters on LUID by @LeoBorcherding in #8863
Studio: add secure keyless localhost and private-LAN API access by @mahiatlinux in #9102
Studio: guard log volume in CI so it cannot regress unnoticed by @danielhanchen in #9665
Studio: take the AMD APU total on Windows from the driver, not the carve-out by @LeoBorcherding in #9314
Security audit: re-approve the unsloth-zoo compiler.py finding after 2026.8.14 by @danielhanchen in #9674
Studio: scope two :has() rules to direct children so they stop walking the whole thread by @danielhanchen in #9669
Remove the temporary ROCm Windows APU probe script by @danielhanchen in #9681
Studio: let workers spawn the killable Hugging Face prefetch child by @oobabooga in #9625
Stop losing the cause of a failed save, and stop dropping 4bit in silence by @danielhanchen in #9678
Studio: don't crash the app when a tool arg isn't a string by @NilayYadav in #9641
fix(studio/pricing): long-context tier starts strictly above the threshold, not at it by @xyzs996 in #9527
Map Qwen3.8-27B to the unsloth repos by @danielhanchen in #9682
Studio: let the download transport be picked in Settings > General by @shimmyshimmer in #9272
Stop the -bf16 notice claiming a request the caller never made by @danielhanchen in #9685
Decide the embedding offload and the multi-GPU split without being asked by @danielhanchen in #9677
Studio: label GPU totals GiB, the unit they are actually in by @danielhanchen in #9570
Studio: refuse an oversized unified-memory image load before the download, not after by @oobabooga in #9629

New Contributors

@divagr18 made their first contribution in #8866
@gokay-ai made their first contribution in #9358
@Siva010 made their first contribution in #9474
@chakshu-dhannawat made their first contribution in #9331
@FrancescoBrunoDev made their first contribution in #9270
@hellopahe made their first contribution in #8953
@kissa01 made their first contribution in #8939
@xyzs996 made their first contribution in #9527

Full Changelog: v0.1.802-beta...v0.1.802-beta

unslothai/unsloth v0.1.802-beta Bug Fixes + Auto compaction + LAN Remote Access on GitHub