unslothai/unsloth v0.1.900-beta on GitHub

We're adding support for decision models, a unified Library for docs and media, document viewer, many Apple Silicon improvements, creation of Skills, and ~4.5× faster image and video generation.

Run and serve Decision Models like Laya (open-source Jev) locally
Skills Editor to create, edit, and delete Skills directly in Desktop
ModelScope model downloading is now here for users who can't use HF
Library + Document Viewer for PDF, Word, Excel, PowerPoint, and chats.
Apple Silicon Improvements including batched serving, structured outputs, and TurboQuant KV cache.
Configure how much of a model each GPU receives.
Faster Image + Video Generation with ~4.5x faster LTX-2.3 clips and 1.7–6.3x faster VAE decoding.

Library + Viewer

View PDF, Word, Excel, and PowerPoint files directly in Unsloth, with links back to their source.
Manage chats, images, videos, everything in our new Library tab.
Improved attachment cards and support for reading more file formats.
Create your own sidebar sections and drag sections to reorder them.

Laya + decision models

Run Laya decision models locally to answer yes/no, multiple-choice, and scoring questions with probabilities.

Enable the Decision API from Settings > API, with model selection and CPU or GPU controls.
Supports the TypeSafe SDK through a Jev-compatible /v1/systemone endpoint.
Runs natively on MLX for Apple Silicon when GPU is selected. [Details](#11603)

Faster image + video generation

LTX-2.3 clips are ~4.5x faster with distilled sampling, compile fixes, and hosted FP8 weights.
Image and video VAE optimizations deliver 1.7–6.3x faster decoding, with total generation speedups depending on the workflow.
MiniMax-H3's first render is up to a minute faster, with 25–29 GiB lower peak memory on A100, B200, and RTX PRO 6000.

Apple Silicon + MLX

Batched serving allows multiple replies to decode at once.
Added grammar-constrained structured outputs through response_format.
Added TurboQuant KV cache and KV cache quantization for sliding-window models such as Gemma 4.
Improved memory estimates and automatic context sizing based on available memory.

Skills + chat improvements

Create, edit, and delete Skills from the Desktop Skills menu.
Improved conversation recall across repeated auto compactions.
Long-running tool calls continue even when they produce no output, with better protection against excessive output consuming memory.
Auto-scroll or manual scroll your conversations
Use ModelScope instead of Hugging Face for model downloads

Training + inference

Faster block-FP8 LoRA training and support for LoRA on compressed-tensors W8A8 checkpoints.
Load Voxtral and Qwen2-Audio through FastModel.
Export saved checkpoints from crashed or cancelled training runs.
Important compatibility fixes for Gemma, newer TRL trainers, and FP8 on RTX 5090 and RTX PRO 6000.

What's Changed

Wait for the model load before reading the unsloth run banner by @danielhanchen in #11679
Bump install.sh / install.ps1 pin to unsloth>=2026.9.11 by @danielhanchen in #11689
Revert "Studio: steady the Images and Video loading spinners, and tidy the progress card" by @shimmyshimmer in #11691
Follow #11635's Diffusers prefetch and #11660's responsive classes in the tests by @danielhanchen in #11680
Give the ROCm bf16 chain test every name _gpu_init imports from device_type by @danielhanchen in #11688
Read the sidebar pin's width only where matchMedia exists, and expect the update card's scaled width by @danielhanchen in #11700
Keep a test's fake pid identity out of the cached owner identity by @danielhanchen in #11708
Expect the unsloth Z-Image-Turbo mirror in the image download queue drive by @danielhanchen in #11711
Classify uv's new --output-format as a value flag in the pip shim by @danielhanchen in #11715
Studio: do not roll back a model switch whose load got no answer by @danielhanchen in #11729
studio: add responses api selection for custom providers by @mahiatlinux in #11354
Count the stripper's whole-buffer work instead of timing it by @danielhanchen in #11735
Studio: keep tracking a new chat's upload when its id commits late by @danielhanchen in #11738
studiobench: compare a code fence one arm scrolled past on its text by @danielhanchen in #11741
Anchor the GRPO MoE aux-loss fail-fast on the assignment, not its expression by @danielhanchen in #11747
Chat UI driver: wait for the Recents thread to load, and let the check fail by @danielhanchen in #11749
Clean-machine trace leg: let git fetch the pinned git+ requirements, nothing else by @danielhanchen in #11751
Studio: add Search Hub to the Images, Video and Audio model pickers by @shimmyshimmer in #11756
Studio: keep the model name and quant whole in the Images and Video header by @shimmyshimmer in #11757
Studio: drag to reorder, add to project and quick download for the image and video galleries by @shimmyshimmer in #11758
Studio: make the Images, Video and Audio settings rail resizable by @shimmyshimmer in #11760
Studio installer: do not run Apple's git shim to ask whether git works by @danielhanchen in #11759
Data settings check: pin the legacy store outcome in the failed-delete check by @danielhanchen in #11763
Update idempotency: count whisper.cpp only when the install has one by @danielhanchen in #11762
Account matrix: cover the gallery move and add-to-project routes by @danielhanchen in #11767
Studio: stop clipping the model description's descenders; follow the resizable rail in the layout contracts by @danielhanchen in #11769
Studio: follow-ups for the media gallery, hub link and rail divider by @shimmyshimmer in #11765
studiobench: do not fail UI parity on repetitions that swap two renderings by @danielhanchen in #11772
MCP HTTP integration test: let the server bind its own port by @danielhanchen in #11777
Export full fine-tunes in 16-bit from the CLI by default by @NilayYadav in #11716
Studio: end Alpaca training samples with the end token by @NilayYadav in #11719
Studio: keep Qwen-Image-2.1 int8 / fp8 compiling on torch 2.12 (CantSplit) by @danielhanchen in #11677
Studio: let an explicit FBCache request engage on Qwen-Image-2.1 and other prefix-KV models by @danielhanchen in #11713
Studio: make the HunyuanImage-2.1 denoiser step capturable as a CUDA graph by @danielhanchen in #11753
Studio: pin, reorder and add to project for Audio history by @shimmyshimmer in #11774
Studio: Show token usage and cache stats on connected provider chats by @NilayYadav in #11717
fix(studio): support preserve thinking for llama.cpp connections by @Imagineer99 in #11706
Studio: stop the max speed tier recompiling on every new prompt length by @danielhanchen in #11731
Layout contract: tie the Train rail clamp to its scroller's padding by @danielhanchen in #11773
Studio: fix broken Chinese, Japanese and emoji text in gpt-oss replies by @NilayYadav in #11720
Studio: Keep Codex reasoning between tool calls on /v1/responses by @NilayYadav in #11726
Studio: Make unsloth start --reasoning on/off take effect by @NilayYadav in #11718
Studio: example prompt per image workflow, remember the last prompt by @shimmyshimmer in #11775
Studio: dock the floating Live monitor beside Run settings by @wasimysaid in #11699
Account matrix: cover the audio gallery move and add-to-project routes by @danielhanchen in #11789
Studio: show the Custom size fields under Output size in unified Edit by @oobabooga in #11690
Fix false ONNX rejection for models with native weights by @wasimysaid in #11695
Fix GGUF vision capability selection by @wasimysaid in #11696
Baseline the two huggingface_hub 1.33.0 / 2.0.0 findings after review by @danielhanchen in #11811
Replace an in-flight chat model load with the latest pick by @wasimysaid in #11697
Prebuilt installers: retry a download the server drops by @danielhanchen in #11816
Load-replacement contract: read the rollback guard, not its one-line spelling by @danielhanchen in #11817
Unsloth Studio (AMD/ROCm): don't turn on cudnn.benchmark for image and video generation by @LeoBorcherding in #11732
Padding-free gate tests: state that UNSLOTH_RETURN_LOGITS is unset by @danielhanchen in #11863
Eval step: restore UNSLOTH_RETURN_LOGITS even when evaluation raises by @danielhanchen in #11865
Studio update launcher tests: give every test a private STUDIO_HOME by @danielhanchen in #11869
Studio: a sandbox read waits out a legacy move's staging window by @danielhanchen in #11876
Agent guides CI: pass optional installer flags only while the installer takes them by @danielhanchen in #11878
fix(studio): keep Stop generating visible while queuing by @Biotrioo in #9123
Studio: give the Windows ROCm torchao stub a version transformers 5 can parse by @danielhanchen in #11640
Give vLLM a valid top_k when fast_inference enables it in the GRPO trainer by @danielhanchen in #11675
Studio: stop unsloth start announcing a switch when --model names the file already loaded by @oobabooga in #11873
Studio: stop holding a passthrough response 5 s when a watcher swallows its cancel by @oobabooga in #11859
Studio: say the installer script is missing instead of showing the PowerShell logo by @oobabooga in #11862
Studio: stop rescanning every model folder on each request that names a model not on disk by @oobabooga in #11872
studiobench: image_upload closes the menu it opened when it gives up by @danielhanchen in #11892
Parallel-isolation guard: exempt the resolver's back-dated staleness precondition by @danielhanchen in #11894
Studio: stop opening a saved chat from re-running a Max Tokens reply by @shimmyshimmer in #11875
Studio: use the browser's timezone for today's date in chat by @NilayYadav in #11852
SentenceTransformer: preserve masks for patched Gemma3 attention by @Etherll in #11866
Studio: initialize new chats before attaching documents by @Imagineer99 in #11838
Studio: stop inflating chat images by re-encoding every one to PNG by @oobabooga in #11889
Studio: fix document search for embedding models other than the default by @NilayYadav in #11853
Studio: train audio datasets on the columns the dataset check found by @NilayYadav in #11850
Studio: follow the JSON format a client asks for on /v1/messages by @NilayYadav in #11854
Studio: let a temporary chat be saved to history by @shimmyshimmer in #11901
Studio Hub: show a running download as Downloading, not as a paused partial by @danielhanchen in #11797
Accept the transformers 5.0 ignore_keys argument in validate_rope by @danielhanchen in #11560
Studio: verify the Diffusers main zip against a pinned SHA-256 by @danielhanchen in #11908
Studio: restyle the save temporary chat popup as a standard dialog by @shimmyshimmer in #11914
Fix text_only 4-bit load and generate for Gemma-4 and other VLM text configs by @danielhanchen in #11684
Studio: add File and View menu items to the macOS desktop app by @shimmyshimmer in #11902
Studio: keep the model's own chat template when the Unsloth one can't render the rows by @oobabooga in #11487
Split the flash-attn prebuilt wheel build across parallel ccache jobs by @oobabooga in #11812
Studio: find older chats by title in chat search by @NilayYadav in #11296
Studio: fix login failures during slow startup by @NilayYadav in #11289
Studio: build Qwen-Image-2.1's token layout once per render instead of every step by @danielhanchen in #11887
Load a VLM through its native image-text class when the repo's auto_map class is untrusted by @danielhanchen in #11613
Unsloth Studio (AMD): floor torch at 2.11 on the gfx103X-all and gfx110X-all families too by @LeoBorcherding in #11834
Studio: read DOCX content controls, tracked insertions and text boxes by @L4XB in #11803
Studio: start settings labels with the setting, not "Show" by @shimmyshimmer in #11924
Studio: align the model selector label and truncate long project names by @shimmyshimmer in #11917
Studio: style the chat scrollbar like Run settings by @shimmyshimmer in #11925
Unsloth Studio (AMD/ROCm): warn about, and refuse, a GPU the installed PyTorch has no kernels for by @LeoBorcherding in #11571
Studio: video auto precision keeps a resident bf16 DiT by @danielhanchen in #11831
Studio: show Theme first in Settings > Appearance by @shimmyshimmer in #11921
Stop the Starling, Yi-chat and LFM2 templates leaking whitespace by @Abhishek-B-R in #11779
Studio: clicking a gallery item keeps the typed prompt by @shimmyshimmer in #11930
Studio: move Read aloud and Edit response into the More menu by @shimmyshimmer in #11920
Studio: make the composer the same width as the chat column by @shimmyshimmer in #11926
UI scale contract: count the save-temporary-chat button among the chat header's 30px controls by @danielhanchen in #11932
studio: allow local imatrix files for gguf export by @mahiatlinux in #11350
Studio: guard export operations when the Hub is unreachable by @Imagineer99 in #11466
Studio: keep a Codex chat working after a tool returns an image on a text-only model by @NilayYadav in #11476
Studio: show example prompts as placeholder hints by @shimmyshimmer in #11931
Studio: keep a chat's attached files when you fork it by @NilayYadav in #11295
Studio: add a multiline send shortcut and spell out what each one does by @shimmyshimmer in #11927
Studio: stream durable chat runs at display frame rate by @shimmyshimmer in #11900
Embed server tests: intercept only the server's own Popen by @danielhanchen in #11934
Studio: build the ConvRot rotation from its definition by @danielhanchen in #11807
Studio: keep all text when adding Word files to a knowledge base by @NilayYadav in #11725
Studio: show when a response was written, in its More menu by @shimmyshimmer in #11928
Studio: load the hosted INT8 pre-quant checkpoints on torchao 0.18 and later by @danielhanchen in #11884
Studio: show context checkpoints apart from the KV cache in the memory estimate by @oobabooga in #11581
Studio: drag to reorder pinned models in the model picker by @shimmyshimmer in #11941
Cast fp16 leftovers to the requested dtype after a text_only pre-quantized load by @danielhanchen in #11692
Studio: translate the inline Read aloud and Edit response settings by @shimmyshimmer in #11933
Accept block_sequence_ids in chunked causal masks on transformers 5.17 by @danielhanchen in #11693
Studio: convert decoded video frames to uint8 on the GPU before the mp4 encode by @danielhanchen in #11879
Studio: fix pinned rows in the model picker by @shimmyshimmer in #11943
Studio: stop rewriting the compile-cache bundle on every warm start, and bound its disk use by @danielhanchen in #11874
Studio: list the Images workflows in the phone sidebar again by @oobabooga in #11936
Studio: reuse unchanged files from an older snapshot instead of re-downloading them by @danielhanchen in #11796
Studio: draw one drop line per gap when dragging sidebar rows by @shimmyshimmer in #11942
Studio: show the full URL while a web fetch waits for approval by @wasimysaid in #11694
Rebuild byte-level tokenizers that transformers v5 loads as LlamaTokenizer by @danielhanchen in #11686
Plan a device map for text_only loads of vision-language models by @danielhanchen in #11584
Installer: stop picking cu126 when a slow NVIDIA driver times out CUDA detection by @danielhanchen in #11916
Studio: skip xFormers when its torch requirement is unmet by @oobabooga in #11847
Studio: add HTTP recording fallback to Dictate by @Etherll in #11075
Studio: propagate the Ollama CUDA runtime to llama-server (replacement for #7563) by @wasimysaid in #11666
Prevent setup-size flash during desktop startup by @wasimysaid in #11910
Studio: return freed host memory to the OS after diffusion and video unload by @danielhanchen in #11795
Walk every sub-config when deciding whether a config carries remote code by @danielhanchen in #11555
Studio: fix vision training with evaluation on when there is no eval split by @NilayYadav in #11851
Studio: stop a compiled Qwen-Image-2.1 render holding 2 GiB of prefix K/V it does not need by @danielhanchen in #11882
Dequantize FP8 weights left raw by text_only and offloaded loads by @danielhanchen in #11841
Studio: video status reports CUDA graphs off when they never engage by @danielhanchen in #11886
fix(studio): show compaction notices for tool-loop checkpoints by @Imagineer99 in #11702
Pin the ARM64 Arrow overlay to a commit before its wheels are signed by @danielhanchen in #11911
Load 4.x remote code and config-only remote code on transformers 5 (Trinity-Large, MiniMax-M3) by @danielhanchen in #11658
Studio: serve the Jev API locally with Laya by @NilayYadav in #11603
Studio: opt-in NVENC for the video mp4 export by @danielhanchen in #11881
Studio: compile the diffusion denoiser on fp16 GPUs when a speed tier is picked (T4 and other pre-Ampere cards) by @danielhanchen in #11899
Studio: whole-model offload points weights back at their host tensors instead of copying them by @danielhanchen in #11764
Studio: pin the kept offload weights on first onload when host RAM allows by @danielhanchen in #11766
fix(studio): count rendered GGUF prompts before shared-KV admission by @Imagineer99 in #10794
Point -bf16 4-bit requests at quantization_config, and warn when a bitsandbytes load quantized nothing by @danielhanchen in #11586
Studio: HunyuanVideo-1.5 padded-text trim on the max speed tier by @danielhanchen in #11824
Studio: release VRAM when an image model is unloaded mid-render by @danielhanchen in #11790
Studio: faster MiniMax-H3 video VAE encode and decode by @danielhanchen in #11801
fix(grpo): use the evaluated model's output head for log-probs by @taking-lying-flat in #11877
Studio: auto step cache (FBCache) only on the max speed tier by @danielhanchen in #11791
Studio: cache text-encoder outputs so repeat prompts skip the encoder by @danielhanchen in #11844
fix(studio): classify MLX requests from the applied chat template override by @Lyxot in #11903
Unsloth Studio (AMD): allow INT8 / FP8 image and video precision without torchao by @danielhanchen in #11631
Studio: count the hosted text encoder and keep the GGUF denoiser resident when only the encoder does not fit by @danielhanchen in #11802
Studio: run an explicit int8 under offload on NVIDIA as torchao-free W8A8 by @danielhanchen in #11712
Studio: apply remembered settings for MLX and safetensors models to API loads by @Lyxot in #11904
Studio: keep bf16 for auto precision on image families that cannot compile by @danielhanchen in #11819
Studio: tile the VAE instead of refusing oversized upscales, and in-app remedies for image refusals by @danielhanchen in #11798
Load and train remote-code multimodal wrappers: Phi-4-reasoning-vision in 4-bit and Nemotron-3-Nano-Omni by @danielhanchen in #11526
Finish a 16bit load of a static per-tensor fp8 checkpoint (Mistral-Small-4) by @danielhanchen in #11531
Studio: static step skip for image models that keeps CUDA graphs by @danielhanchen in #11737
Studio: static step skip for video generation by @danielhanchen in #11748
Studio: Qwen-Image-2.1 RoPE in real arithmetic inside the compiled blocks by @danielhanchen in #11888
Studio: Library page for files, media and fine-tunes by @shimmyshimmer in #11770
Resolve remote-code model classes, reach per-expert submodules with LoRA, run the root dispatch hook on the embedding's device (Nemotron-Labs-Teacher) by @danielhanchen in #11543
Studio: Library settings, storage and sortable list columns by @shimmyshimmer in #11771
Studio: open Images and Video previews in the Library viewer by @shimmyshimmer in #11776
Format the two files #11526 left off the formatter's fixed point by @danielhanchen in #11955
Load and fine-tune LongCat-Flash-Lite-Sparse on transformers' longcat_flash by @danielhanchen in #11620
Hand a composition with no forward of its own to its thinker (Qwen3-Omni) by @danielhanchen in #11523
fix(studio): retain GGUF quants across cache-folder switches by @wasimysaid in #11938
Keep the training state intact across a standalone evaluate() / predict() by @danielhanchen in #11860
fix(studio): recover from a dead Metal GPU queue instead of failing every later request by @Lyxot in #11383
Pin the compiled mask wrapper only when it is compiled by @danielhanchen in #11957
Load NVIDIA ModelOpt FP8 checkpoints through the transformers fp8 quantizer by @danielhanchen in #11592
Studio: run Qwen-Image-2.1 int8/fp8 on 24 GB cards with the transformer resident and the text encoder streamed by @danielhanchen in #11883
Divide by gradient accumulation for forwards that take **kwargs but return a mean loss by @danielhanchen in #11898
Load 4.x-era configs that transformers 5 strict validation rejects (Llama 4 attn_temperature_tuning) by @danielhanchen in #11836
Studio: minimal OS sandbox for Python and Terminal tools on Linux and macOS by @oobabooga in #11209
Studio: add Windows MXC Preview sandboxing by @Etherll in #11390
Studio: whole-model NVFP4 for the video families with hosted pre-quantized denoisers by @danielhanchen in #10729
Studio: per-layer NVFP4 image policies, flashinfer FP4 backend and a gated auto row by @danielhanchen in #10730
Studio: NVFP4 flashinfer backend kernel items (device guard, persistent barrier, bias path, cached dispatch) by @danielhanchen in #10731
Studio: keep inline HTML text in its line when parsing an upload by @L4XB in #11818
Studio: read text, Markdown and HTML uploads in the encoding they were written in by @L4XB in #11896
Fix PEFT base export targets silently exporting the base model by @tweekli in #11781
Studio: install flashinfer on demand for NVFP4 without moving torch by @danielhanchen in #11730
Studio: steady pinned model drop line and pill-shaped project drop highlight by @shimmyshimmer in #11986
Studio: compile the VAE decode for DiT families, from the NVFP4 time budget pass by @danielhanchen in #10889
Studio: Library grid Sort menu, tidier toolbar menus and card icons by @shimmyshimmer in #11978
Studio: cap context checkpoints by host RAM for sliding-window and SSM models by @Lyxot in #11918
Studio: sort projects in the sidebar and show pinned chats once by @shimmyshimmer in #11988
Studio tests: read UI labels from the en catalog, and follow the response-details action into MessageMenuTime by @danielhanchen in #11949
Studio: open model row tooltips from the name only by @shimmyshimmer in #11985
Studio: pin fine-tuned models in the model picker by @shimmyshimmer in #11984
CI: only cancel superseded pull request runs, never dispatches or schedules by @danielhanchen in #11977
CI: fix path filters that miss dependencies, and narrow three that are too broad by @danielhanchen in #11989
Studio Playwright helpers: named per-step budgets, fail-fast step report, condition waits by @danielhanchen in #11979
Studio Playwright chat_ui: condition waits, per-step budgets, 25 s less in the theme step by @danielhanchen in #11983
Studio Playwright: condition waits and per-step budgets in extra_ui and update_banner_layout by @danielhanchen in #11990
Studio Playwright: condition waits and per-step budgets in model_config and memory_estimate by @danielhanchen in #11987
Studio Playwright: condition waits and per-step budgets in loaded_models_indicator and ui_font_scale by @danielhanchen in #11992
Studio Playwright: condition waits and step budgets in thread_scoped_settings, mcp_arguments, chat_width by @danielhanchen in #11991
Export: add save_pretrained_openvino and push_to_hub_openvino support by @goodmai in #11907
Studio: line sidebar section headers up with row content by @shimmyshimmer in #11996
CLI: keep a reply's trailing '<' or '<th' once the stream has ended by @breken-ai in #11893
Studio: drop Move up / Move down from sidebar row menus by @danielhanchen in #12000
Studio: Help menu items and a Go menu for desktop Help search by @shimmyshimmer in #12002
Studio: size Library gallery items without resolving each file by @danielhanchen in #12003
Fix left-padded Online DPO scoring during training by @taking-lying-flat in #11885
feat(studio): ModelScope as a model source and a custom Hugging Face endpoint by @Lyxot in #11761
Keep the Kaggle GPU harness tests independent of the caller's CUDA_VISIBLE_DEVICES by @danielhanchen in #12010
OpenVINO export: trust remote code only for a remote-code model, and decode the bounds probe as UTF-8 by @danielhanchen in #12012
Skip test_wait_for_settled when playwright.sync_api is only a stub by @danielhanchen in #12013
feat(install): fall back to CERNET and npmmirror when package hosts are blocked or slow by @Lyxot in #11786
Skip Unsloth's generated compile cache in the exec-literal lint by @danielhanchen in #12014
tests: make the ROCm install suite pass on Windows, macOS and arm64 runners by @danielhanchen in #12005
Studio setup: read the amd-smi index-space line without head -n 1 by @danielhanchen in #12004
Add longcat_flash_lsa to the fused-MoE conversion snapshot by @danielhanchen in #12018
Record only the test thread's sleeps as Deep Research retry backoff by @danielhanchen in #12019
Studio CLI: stop unsloth run re-exec'ing itself forever when the Studio venv is a symlink by @danielhanchen in #11788
Studio: do not reapply ROCR_VISIBLE_DEVICES when picking the AMD card in setup.sh and the llama.cpp prebuilt probe by @danielhanchen in #11965
Studio: run image and video denoises on one render thread so cuDNN caches are reused by @danielhanchen in #11843
Studio: return an error when embeddings dimensions can't be honored by @NilayYadav in #11968
Studio: keep a 0 label and blank a NaN cell when mapping columns to chat roles by @breken-ai in #11895
Studio: support web_search_20260209 on /v1/messages by @NilayYadav in #11963
Studio: keep a working GPU when one probe fails, and name the card behind no_gpu by @danielhanchen in #11944
Studio: download only one copy of a GGUF quant by @NilayYadav in #11966
Give the real-host NVIDIA probe test a budget that a busy driver can meet by @danielhanchen in #12026
Unsloth Studio installer (AMD/Linux): send RDNA 4 cards (RX 9000, R9700) to AMD's gfx120X wheels by @danielhanchen in #11935
Studio: keep a CSV seed's values as written when dropping its index column by @L4XB in #11945
Studio media viewer: Zoom to fit at the bottom of the scale menu by @shimmyshimmer in #12028
Studio: shadow the dark mode composer so it stands off the chat by @shimmyshimmer in #12023
Studio: name the Git Bash MXC incompatibility and stop re-probing it by @danielhanchen in #12021
Keep more VRAM headroom on Windows CUDA, and say when a hand-set context does not fit by @danielhanchen in #11368
Unsloth Studio (AMD): keep export off a GPU PyTorch has no kernels for, like the iGPU by @oobabooga in #11946
Read the hot-path I/O cost at its steady minimum across repeats by @danielhanchen in #12029
Unsloth Studio: a companion fetch with no denoiser no longer blocks deleting its base by @LeoBorcherding in #11828
Studio: stop Python tool network calls from skipping the host allowlist by @oobabooga in #11172
Studio: center collapsed sidebar icons and tighten the rail by @shimmyshimmer in #12031
Unsloth Studio: list a GGUF with no header metadata under On Device on the Images page by @LeoBorcherding in #11830
Keep the conversion backfill's donor stub off the transformers package by @danielhanchen in #12034
Unsloth Desktop: show Stopping… after Stop so a pending cancel doesn't look ignored by @LeoBorcherding in #11976
Studio: stop edit_file writing a compacted-argument placeholder into files by @oobabooga in #11950
Studio: decode Wan video in fp16 with channels_last_3d convs by @danielhanchen in #11999
Unsloth Studio: refuse picking a hosted FP8/INT8 checkpoint repo as a pipeline before it downloads by @LeoBorcherding in #11829
Studio: stop the second prompt length recompiling Qwen-Image on the max tier and with int8 by @danielhanchen in #11842
Studio: size the image memory plan at the dtype the pipeline loads in (SDXL resident on 24 GB) by @danielhanchen in #11922
Studio: stop MiniMax-H3 recompiling on the second caption and the first i2v by @danielhanchen in #11880
Unsloth Studio / Desktop: log what an image load and each generation actually resolved to by @LeoBorcherding in #11994
Unsloth Studio: stop offering "Continue" on a base repo that only holds a GGUF's text encoder and VAE by @LeoBorcherding in #11644
Apply SFTConfig.router_aux_loss_coef to MoE models that cache it at init by @danielhanchen in #12006
Load only the language model for text_only on repo-code composites by @danielhanchen in #11861
Put back the unsloth_zoo modules the device map opt-in tests stub by @danielhanchen in #12038
Studio: pass tool_result is_error through to the model on /v1/messages by @NilayYadav in #11962
SentenceTransformer: add opt-in FP32 merged-pair ranking loss by @Etherll in #11867
Studio: opt-in Hadamard rotation for Qwen-Image-2.1's int8 transformer so it matches bf16 by @danielhanchen in #11835
Grouped-linear LoRA for DeepSeek-V4, remote-code shims for Step-3.7, and a real message for Mistral-format checkpoints by @danielhanchen in #11528
Studio: fall back to native when SageAttention does not run on this GPU by @danielhanchen in #11997
Keep flash attention off sub-models that do not support it (LFM2.5-VL SigLIP2 tower) by @danielhanchen in #11959
Studio: stop trading the denoiser's quantisation away to pay for offload by @danielhanchen in #11558
Unsloth Studio: report the VAE decode on the image progress bar by @LeoBorcherding in #11740
Translate the Library toolbar Sort menu in every locale by @danielhanchen in #12045
Read the partial safetensors delete-menu guard by operator, not verbatim by @danielhanchen in #12042
Resolve the macOS app menu's chords as a Mac in its test by @danielhanchen in #12046
Advertise NVFP4 diffusion on the text encoder driver's mocked host by @danielhanchen in #12047
Keep peft's is_torchao_available cache API through the stale-torchao patch by @danielhanchen in #12049
Run the formatter fixed-point guard's batches side by side by @danielhanchen in #12050
Let the macOS tab sampler ride out a navigation still in flight after login by @danielhanchen in #12053
Studio: let Deep Research finish a turn handed off from a chat generation by @MohammadHijjawi97 in #11923
Studio: custom sidebar sections, section menus and drag to reorder by @shimmyshimmer in #12016
Studio: link folders when creating a project by @shimmyshimmer in #12057
Studio: open the user menu Help submenu upward by @shimmyshimmer in #12032
Skip chordless menu items in the native chord collision check by @danielhanchen in #12063
Give the health wait's working-child tests room for the worker to start by @danielhanchen in #12065
Count #12016's section header among the sidebar's scaled 30px rows by @danielhanchen in #12066
Turn off Dr GRPO reward scaling under TRL's "group" default by @vineethsaivs in #11951
Studio: keep the sidebar menu shadow in light mode by @shimmyshimmer in #12062
Reset the permission step's storage at the start of the next document by @danielhanchen in #12073
Let the Projects section stand in for its row in the macOS tab walk by @danielhanchen in #12074
Load #11526's text-core refusal in the save_method routing harness by @danielhanchen in #12077
Studio: read personalization only once a first sign-in has changed its password by @danielhanchen in #12071
Studio: train a SQuAD answer's text, not the answers dict, when mapping columns to chat roles by @breken-ai in #12056
Back off and retry the update banner navigation on ERR_NO_BUFFER_SPACE by @danielhanchen in #12081
Keep TRL's own RL config defaults and clamp preference max_length by @danielhanchen in #12069
Unsloth Studio (AMD/Windows): report an iGPU's used VRAM when a discrete card sits beside it by @LeoBorcherding in #11871
CI: CodeQL advanced setup that analyses only the languages a PR touches by @danielhanchen in #11998
Give the killed formatter grandchild as long to die as it had to start by @danielhanchen in #12086
Studio: rework Appearance settings and add flavor color themes by @shimmyshimmer in #12030
Studio: quieter settings headings and one section gap on every page by @shimmyshimmer in #12064
Keep the padding-free column test off the datasets numpy formatter by @danielhanchen in #12082
Studio: put Reapply next to Generate on the image and video pages by @LeoBorcherding in #11974
Unsloth Studio / Desktop: keep the Images Cancel load pill off the model panel divider by @LeoBorcherding in #12079
Studio: keep run duration on Mac when training with an eval set by @NilayYadav in #11967
fix: handle strided cross entropy inputs by @MrCapricornLiu in #10713
Stop unsloth chat and unsloth inference switching a loaded GGUF to a different quant by @NilayYadav in #11855
Studio: tell Safari which account the Studio password belongs to by @NilayYadav in #11858
Studio: find models in a Hugging Face cache folder added as a location by @NilayYadav in #11970
Unsloth Studio installer (AMD/Windows): don't mistake ZLUDA for an NVIDIA GPU by @LeoBorcherding in #11736
Studio: follow redirect pages when reading a web page by @NilayYadav in #11856
Studio: don't stop long running tool calls that print nothing by @NilayYadav in #11969
Studio: stop the dense-quant probes pinning a CUDA context on every card of a multi-GPU host by @LeoBorcherding in #11954
Studio: reject echo, suffix and best_of on /v1/completions by @NilayYadav in #11964
Fix the prebuilt wheel publish step and shorten its release notes by @danielhanchen in #12068
Studio: stop generating when the client disconnects on a non-streaming request by @NilayYadav in #11961
Studio: decode the SDXL VAE in fp16 on fp16 GPUs by @danielhanchen in #12036
Studio: honor response_format on the MLX backend with grammar-constrained decoding by @Lyxot in #10180
Studio: document viewer for PDF, Word, Excel and PowerPoint, with origin links in Library by @shimmyshimmer in #12001
Studio: chat attachment cards, chips and the Library viewer by @shimmyshimmer in #12017
Studio: keep an eagerly decoded image VAE contiguous on NVIDIA by @danielhanchen in #12035
Studio: skip the cuDNN benchmark search for the MiniMax-H3 audio VAE on A100 / B200 / RTX PRO 6000 (first render up to a minute faster, 25-29 GiB lower peak) by @danielhanchen in #12040
Studio: keep a chat's start date in the system prompt, note a new date on the latest user turn by @danielhanchen in #12096
Keep Llama 3.2 Vision off flash attention (vision and cross attention have no is_causal) by @danielhanchen in #12033
Load the repo AutoProcessor for AutoModel-only repo-code VLMs by @danielhanchen in #12037
Unsloth Studio / Desktop: let the GPUs picker say how much of the model each card gets by @LeoBorcherding in #12015
Studio: regionally compile Lumina-2 and HiDream-I1, and re-decide their auto precision by measurement by @danielhanchen in #12039
Studio: refuse a hand-set Metal context only past the GPU wired limit by @Lyxot in #10804
Studio: keep a pinned context as a request limit instead of refusing KV cache quantization by @Lyxot in #11084
Studio: let full-scope keyless callers auto-switch models, and explain the refusal elsewhere by @Lyxot in #11180
Upload all GGUF files in one commit so create_pr opens one pull request by @NilayYadav in #11857
Studio: return an error when a model can't use the tools sent to it by @NilayYadav in #11960
Read settings.py as utf-8 in the palette filter test by @danielhanchen in #12105
Read settings.py as UTF-8 in the unknown-palette test by @danielhanchen in #12101
Carry the resolved attention implementation to nested configs a remote config baked flash attention into (Nemotron 3 Nano Omni) by @danielhanchen in #12099
Studio: share MLX VLM prompt-cache snapshot buffers and replay exact prompts by @Lyxot in #11659
GRPO: default to TRL's dapo loss with beta 0, cap CISPO weights at 5.0 by @danielhanchen in #12088
Patch the TRL trainers that moved to trl.experimental (ORPO, CPO, Online DPO, GKD, ...) by @danielhanchen in #12097
Studio: quantize the KV cache of sliding-window MLX models such as Gemma 4 by @Lyxot in #11082
Studio: cache and coalesce nvidia-smi reads in the backend by @danielhanchen in #11995
Unsloth Studio installer (AMD/Linux): explain why a newer ROCm gets ROCm 7.2 PyTorch by @LeoBorcherding in #11651
Treat a reaped grandchild as dead in the formatter timeout test by @danielhanchen in #12107
Trim comments in the DeepSeek-V4 grouped LoRA, remote-code shims and Mistral-format loader code by @danielhanchen in #12089
Keep ORPO / CPO rows within max_length on TRL 0.29+ by @danielhanchen in #12115
Keep the Triton MoE grouped GEMM in compiled graphs and index weights past 2^31 elements by @danielhanchen in #12114
Studio: read more chat attachment formats, and hand the rest to the python tool by @Lyxot in #11379
Unsloth Desktop: Create, Edit, and Delete Skills from the Skills Menu by @LeoBorcherding in #11800
Studio: show the Hub error for an unreadable GGUF repo instead of routing it to Transformers by @danielhanchen in #12117
Count the MLX grammar engine slot in the Apple Silicon step totals by @danielhanchen in #12135
Pass gradients through compressed-tensors activation quantization so W8A8 checkpoints train with LoRA by @danielhanchen in #11585
Studio: keep the last full-attention layer unquantized in the MLX KV cache by @Lyxot in #11083
Gate the Mllama CUDA forward test on has_real_cuda by @danielhanchen in #12143
Batched serving on the MLX path: several replies decoding at once by @Lyxot in #10310
Point dsh at Unsloth through a --patch overlay instead of settings.yaml by @danielhanchen in #12145
fix(studio): mark /api reads no-store so an idle desktop app stops rewriting its disk cache by @alkinun in #12148
Wait for every diffusion run a test started before undoing its runs dir by @danielhanchen in #12149
Studio: move Managed accounts to the Accounts tab and drop Mark as unread from chat menus by @shimmyshimmer in #12120
Studio: keep checkpoint compaction under --disable-tools by @danielhanchen in #12119
Studio: only offer Agent Skills when Code is on by @danielhanchen in #12118
Settle the permission step's reloads on the pill instead of networkidle by @danielhanchen in #12153
Speed up block-FP8 LoRA training: run FP8 linears eagerly, 8 warps for 128-row GEMM tiles by @danielhanchen in #12027
Fall back from FBGEMM for rowwise FP8 on GPUs it has no kernel for (RTX PRO 6000 / 5090) by @danielhanchen in #12098
Studio: TurboQuant KV cache option for MLX inference by @Lyxot in #11170
Fix rowwise FP8 scale axes in fused LoRA backward by @taking-lying-flat in #11799
Unsloth Studio Installer: ask Python for a path identity before giving up on an exact one by @danielhanchen in #11104
Ask Python for the process image table when the native helper is unavailable by @danielhanchen in #11115
Refresh a rewritten shortcut's icon where the shell cannot define the type by @danielhanchen in #11116
Recover CUDA compute capabilities without emitting a P/Invoke type by @danielhanchen in #11173
Remove the reflection-emit apparatus and all four emitted types by @danielhanchen in #11193
Add a Windows probe for code integrity blocks, and audit bundle signatures in CI by @danielhanchen in #10408
Windows installers find an NVIDIA GPU on the PCI bus, and say which CUDA it can use by @danielhanchen in #11166
Fix Gemma2 padding masks during batched cached decoding by @taking-lying-flat in #12008
Gemma2: use flash_attn_with_kvcache for cached decoding by @danielhanchen in #12112
Fix Gemma and Gemma2 embeddings scaled twice on transformers 5.4+ by @danielhanchen in #12116
Keep scan_packages.py from tripping Bitdefender's Python stealer signature by @danielhanchen in #12169
Studio: show when a prompt was sent while hovering it by @shimmyshimmer in #12170
Studio: add a Scroll while generating setting (Auto-scroll or Manual) by @shimmyshimmer in #12110
Studio: make compile knobs reach the render thread on torch 2.12+ by @danielhanchen in #12075
Studio: stop a hung system node or npm from stalling setup by @oobabooga in #12165
Studio: price MLX loads in the memory panel and fit an unpinned context to available memory by @Lyxot in #10287
Gemma2: keep softcapping attention under int32 indexing and fall back to eager if compile fails by @danielhanchen in #12154
Fix training with accelerate 1.15 on torch without a distributed backend (AMD Windows ROCm) by @danielhanchen in #12162
Studio: stop runaway tool output from using up memory by @NilayYadav in #11723
Studio: reset the download progress bar when a retry restarts the file by @NilayYadav in #11593
Studio: Select all only picks the models the search shows by @NilayYadav in #11485
Studio: ignore an SSLKEYLOGFILE the process cannot write instead of failing every HTTPS client by @danielhanchen in #12166
Tests: stop Windows tests tripping Bitdefender and the 16-bit application dialog on real machines by @danielhanchen in #12167
Studio: Chats library in the Library by @shimmyshimmer in #12122
Studio: round hover for the Settings close button by @shimmyshimmer in #12172
unsloth start opencode: size the output limit to the context and add --max-tokens by @shimmyshimmer in #12111
SAC probe: verify Studio identity before sending a password by @danielhanchen in #12176
Attention resolver: respect a declared _supports_sdpa = False, and skip flash_attention_2 when a class's compatible flash kernels exclude it (MiMo-V2-Flash) by @danielhanchen in #12147
Keep Nemotron-H mixer.out_proj unquantized under a caller's BitsAndBytesConfig by @danielhanchen in #12131
Phi-4-reasoning-vision: give remote multimodal prep an indexable cache view on transformers 5 by @danielhanchen in #12123
Studio: LTX-2.3 about 4.5x faster per clip (unguided distilled sampling, compile fixes, hosted FP8) by @danielhanchen in #12067
Let a loaded Kimi K2.5 / K2.7 processor take processor(text=..., images=...) by @danielhanchen in #12126
Build the native image processor at defaults when a VLM repo has no preprocessor_config.json by @danielhanchen in #12125
Refuse K-EXAONE 2.0 on a transformers that ignores its config by @danielhanchen in #12128
Studio: list at most the 12 most recently active projects and sections in Move to by @shimmyshimmer in #12155
Turn off the MoE aux loss for dense models that carry a router config by @danielhanchen in #12076
Studio: keep conversation recall working across repeated compactions by @danielhanchen in #12174
Studio: never ask for approval to run search_conversation by @danielhanchen in #12175
Add a CI gate for code shapes that heuristic antivirus scanners quarantine by @danielhanchen in #12178
Dequantize block-FP8 weights with a ragged last block on 16-bit loads (GLM-5.3) by @danielhanchen in #12133
Rebuild CohereTokenizer from tokenizer.json when transformers v5 changes its ids by @danielhanchen in #12130
Keep flash attention off towers the Auto classes do not register by @danielhanchen in #12129
Keep Linear layers an FP8 checkpoint stores in bf16 unconverted by @danielhanchen in #12124
Load speech-to-text models (Voxtral, Qwen2-Audio) through FastModel by @danielhanchen in #12127
Narrow Kimi-K3's zero-padded KDA A_log to num_heads so the fla backward runs by @danielhanchen in #12132
Repair chat templates that always append the generation prompt, also on FastModel loads by @danielhanchen in #12139
Route GKD distillation through the chunked generalized JSD by @danielhanchen in #12136
Studio: list HF cache models written without symlinks in the local inventory by @danielhanchen in #12156
Studio: read public Hub repos without a saved token the Hub rejects by @danielhanchen in #12158
Studio: list a crashed or cancelled run's saved checkpoints on the Export page by @NilayYadav in #11600
Studio: pick the right model size when the file names are lowercase by @NilayYadav in #11722
Refresh WSL shortcut icons through Python instead of an emitted native stub by @danielhanchen in #12184
Read the gradient checkpointing precondition off transformers' own method by @danielhanchen in #12187
Studio: Triton-fused VAE norms, caches and attention for every image and video VAE (1.7x to 6.3x decode) by @danielhanchen in #12078
Studio: fix Decision API validation and TypeSafe SDK metadata by @wasimysaid in #12186
Studio: grant MXC Tier 3 read access to runtime folders once, not per launch by @danielhanchen in #12121
Studio: tell a rejected Hugging Face token apart from an unreachable Hub by @danielhanchen in #12159
Studio: run the isolated Windows Terminal on cmd.exe with stock git when Git Bash cannot start in MXC by @danielhanchen in #12164
Studio: load a downloaded GGUF from disk when the Hub cannot be read or refuses it by @danielhanchen in #12161
Join the npm scanner's credential path markers from pieces by @danielhanchen in #12181
Pass the MXC host-prep script as plain text instead of an encoded command by @danielhanchen in #12182
Count only the retry loop's own sleeps in the JSON fallback backoff test by @danielhanchen in #12189
Store the Studio sandbox credential path names in pieces by @danielhanchen in #12188
Name utf-8 on the SSLKEYLOGFILE writability probe by @danielhanchen in #12190
Studio: fused int8 MLP kernels and real-arithmetic RoPE for DiT denoisers (FLUX.1 11% less GPU time per step) by @danielhanchen in #12083
Studio: capability tags in the model picker match the vision pill and get their own colours by @danielhanchen in #12192
Studio: fix inductor CantSplit on torch 2.12/2.13 and speed up the Qwen-Image-2.1 VAE by @danielhanchen in #12059
Load block-FP8 checkpoints in 4-bit (NF4) when load_in_4bit=True is passed by @danielhanchen in #12146
Studio: compile FLUX.1 dynamic, CUDA-graph the SDXL U-Net, decode one-frame Qwen-Image latents in 2D by @danielhanchen in #12060
Studio: compile VAEs by repeated block, one H3 graph per first render, vectorise the HunyuanVideo-1.5 VAE mask by @danielhanchen in #12061
Unsloth Studio: stop a Mac browser hiding GPU-only models from a remote CUDA / ROCm / Intel server by @danielhanchen in #8833
Studio: return no-op instead of 500 for an out-of-range scan folder id by @danielhanchen in #8397
Do not reinstall llm-compressor when it is already installed by @danielhanchen in #6806
Fix Studio launcher repair for missing install id by @danielhanchen in #6933
Load Mistral-format checkpoints (params.json only) through transformers' Mistral4 (Mistral-Large-3) by @danielhanchen in #12144
Studio: name the MCP server while the tool call is still streaming by @danielhanchen in #9212
Stop gpt-oss generation at the harmony tool call token by @danielhanchen in #11449
Studio: vendor laya 0.3.5 and hold its weights in float16 by @danielhanchen in #12202
Unpack the mirrored uv wheel with python3 -m zipfile instead of an inline one-liner by @danielhanchen in #12194
Keep xFormers attention causal when the decoder is called without a mask by @danielhanchen in #12199
GKD: right-align left-padded rows before the student forward by @danielhanchen in #12200

New Contributors

@L4XB made their first contribution in #11803
@Abhishek-B-R made their first contribution in #11779
@tweekli made their first contribution in #11781
@goodmai made their first contribution in #11907
@breken-ai made their first contribution in #11893
@MohammadHijjawi97 made their first contribution in #11923

Full Changelog: v0.1.815-beta...v0.1.900-beta

unslothai/unsloth v0.1.900-beta Laya Decision Models + Library on GitHub