unslothai/unsloth v0.1.906-beta on GitHub

Models can now use the browser in Unsloth. Turn on Browser use and your model can open pages, read them, click, type and scroll while you watch. Qwen-Image-2.1-Turbo is here too, making images in just 8 steps. Also included: text-diffusion LLM training, faster MoE GGUFs and speculative decoding on Mac.

Highlights

Browser use. Local models can take control of the Unsloth Browser
Qwen-Image-2.1-Turbo: images in 8 steps (2.5x faster). See Guide
Train text-diffusion LLMs like DiffusionGemma, LLaDA and Dream
Up to 3x faster MoE GGUFs that are bigger than your VRAM
Speculative decoding on Mac for up to 2.8x faster MLX replies

Browser use

Turn on Browser use from the + menu in chat, then ask your model to look something up or use a website.
Models can open pages, read them, click, type, pick options and scroll while you watch.
The model browses in its own tab beside the chat, with no access to your other tabs.
Browser use works with local models and API providers, including small text-only models.
Unsloth asks you before the model signs in, posts a form or presses a button that buys, sends or deletes.
The model never fills in passwords, card numbers or one-time codes. Those are yours to enter.
Choose which sites models can open, and their access to your browsing history and downloads, in Settings > Browser.

Image generation

Run Qwen-Image-2.1-Turbo, the official 8-step version of Qwen-Image-2.1, from the Images page, with a 4.2 GB Q4 GGUF. GGUF
Export the image LoRAs you train in Unsloth as files you can reuse or share, and load image LoRAs from your own folders.
ComfyUI models now use the text encoders already in your ComfyUI folder instead of downloading them again.

Training

Train text-diffusion LLMs (DiffusionGemma, LLaDA, LLaDA2, Dream, Nemotron and SDAR) with the new DiffusionTrainer.
Train with LoRA+ by setting loraplus_lr_ratio, and fine-tune GPTQ models with LoRA.
GRPO now trains on video prompts and tool calls with TRL 1.15, and AsyncGRPOTrainer can train and generate on one GPU.
Fine-tuning on tool-calling datasets now includes each row's tool definitions, and works on multi-turn agent traces.
On the Train page, set Offload layers with a VRAM budget (NVIDIA and AMD), and continue LoRA training from a saved adapter.
Train Laya and Clef decision models on Mac. Clef can also train on images.

Performance

MoE GGUFs bigger than your VRAM now run up to 3x faster on NVIDIA and AMD GPUs.
Training with sample packing is now up to 8x faster when flash-attn and xformers are not installed (NVIDIA, PyTorch 2.12+).
Speculative decoding makes MLX replies on Mac up to 2.8x faster, with the same output.
MLX speculative decoding sets its draft depth automatically, and supports MTP, DFlash, DSpark and EAGLE-3 drafters.
Typing and scrolling stay smooth in long chats, and with hundreds of chats in the sidebar.

Chat + Desktop

React code blocks (JSX and TSX) in a reply now run as live previews in the Browser.
Deep Research can use your MCP tools as sources and cite them, and retrying a failed run resumes from the steps it already finished.
Unsloth now reads AGENTS.md from your Unsloth home and project folders into chat instructions.
In Desktop, a project can now run its tools in an existing folder on your computer.
Turn on Read replies automatically in Settings > Voice to hear each reply aloud when it finishes.

Models + settings

Use the optional vLLM engine on AMD GPUs, on Linux and through WSL on Windows.
Run DeepSeek-OCR and DeepSeek-OCR-2.
Filter Hub models by parameter count, context length and organization.
Turn on the detailed Live monitor to see GPU load, temperature and power.
On multi-user installs, limit which models each managed account can use.
Install the Unsloth web UI as an app from a supported browser.

What's Changed

Bump install.sh / install.ps1 pins to unsloth>=2026.10.3 by @danielhanchen in #13083
Keep loss-only evaluation on the fused CE path instead of materializing logits by @danielhanchen in #13065
Studio: int8 ConvRot Qwen-Image-2.1 text encoder on Apple Silicon, and release text encoders during denoise on unified memory by @danielhanchen in #13026
Save models loaded from pre-quantized bitsandbytes checkpoints on transformers 5 by @danielhanchen in #13054
Studio: re-measure annotate marks after a tab zoom by @shimmyshimmer in #13082
Studio: install the released Diffusers 0.41.0 instead of the pinned main commit by @danielhanchen in #13045
Studio: restart llama-server after a lost GPU device instead of locking the chat by @danielhanchen in #13044
Studio: keep CUDA 12 llama.cpp prebuilts off drivers older than CUDA 12.4 by @danielhanchen in #13031
Studio: don't take a port another process holds on 0.0.0.0 by @SulimanAbdulrazzaq in #12011
Studio: load community SDXL single-file GGUFs on the Images page by @danielhanchen in #13035
Studio: fix Qwen-Image-2.1 input-image encode on Apple Silicon (MPS frame-axis pad returns wrong data) by @Lyxot in #13028
Dequantize pre-quantized bitsandbytes checkpoints for full finetuning by @danielhanchen in #13079
Add loraplus_lr_ratio for LoRA+ training by @danielhanchen in #13062
MLX: train decision models on Apple Silicon by @Lyxot in #13016
Unsloth Studio / Desktop: Block Swap settings, a VRAM budget, and a live layer panel on the training page by @LeoBorcherding in #12996
Unsloth Studio / Desktop: per-GPU Block Swap budget and layer panel for multi-GPU training by @LeoBorcherding in #12998
Studio: keep typing and streaming fast in long chats by @danielhanchen in #13059
Reject Activated LoRA adapters instead of training them as plain LoRA by @danielhanchen in #13089
Studio: provide Git lazily for FP8/FP4 export instead of at install time by @danielhanchen in #13084
Studio: let Gemma 4 GGUF models think after a tool call by @NilayYadav in #13097
GGUF adapter export and llama.cpp installer: use unslothai/llama.cpp releases only, no git by @danielhanchen in #13056
Windows installer: stop installing Git and the VS toolchain without asking by @LeoBorcherding in #12971
Studio: keep table columns in attached Word and HTML files by @NilayYadav in #13105
Studio: stop link URLs from crowding out long page text by @NilayYadav in #13102
Studio: keep Think off working on Claude Opus 5.5 and Sonnet 5.5 by @NilayYadav in #13104
Studio: raise the llama-server micro-batch to 2048 when MoE experts spill to RAM by @danielhanchen in #12950
Studio: let imported chats with image links answer follow-up questions by @NilayYadav in #13099
Studio: hold message hover still while a wheel scroll moves the thread by @danielhanchen in #13077
Studio: fill the context bar for Ollama connections by @NilayYadav in #13106
Studio: read news articles whose pages carry megabytes of inline code by @NilayYadav in #13100
Studio: open Images workflows from More when Images is unpinned by @NilayYadav in #13101
Studio: let Deep Research search MCP tools as extra sources by @NilayYadav in #13103
Studio: let Gemma 4 read web search and tool results on safetensors by @NilayYadav in #13096
Studio: reopen a chat on the branch it was left on by @oobabooga in #13109
fix(chat): import branch threads parents-before-children on refresh by @CelesteHeartsong in #13113
Studio: show Gemini's thinking in chat by @NilayYadav in #13107
Studio: show the context meter for active project chats by @lorenzozanee in #12605
Fix returned logit scaling and softcapping when labels are provided by @taking-lying-flat in #13069
Studio: head a decision score with its most likely level by @LeoBorcherding in #13111
Keep slow tokenizers transformers cannot convert instead of crashing the load by @danielhanchen in #13073
Turn torch.compile off on GPUs older than Volta by @danielhanchen in #13085
Keep paged bitsandbytes optimizer state paged on checkpoint resume by @danielhanchen in #13092
Fix LLaVA image token count when processor_config lacks num_additional_image_tokens by @danielhanchen in #13091
Let GPTQ checkpoints train and generate with LoRA by @danielhanchen in #13068
Fix PPOTrainer with an Unsloth policy (gradient checkpointing, generate, RoPE on value and reward models) by @danielhanchen in #13067
Fix beam search (num_beams > 1) on the fast decode path by @danielhanchen in #13053
Fix TRL PPOTrainer under Unsloth: rollout crash, sampling filters and left padding by @danielhanchen in #13108
Load Whisper in FastModel without auto_model by @danielhanchen in #13080
Repair RoPE for plain transformers loads after an Unsloth load (reward models in Online DPO / GRPO) by @danielhanchen in #13055
Point a 4-bit load that does not fit the GPU at offload_layers = "auto" by @danielhanchen in #13052
Make precomputed inputs_embeds train under gradient checkpointing by @danielhanchen in #13088
Right-pad encoder-decoder tokenizers and load SeamlessM4T without auto_model by @danielhanchen in #13066
SyntheticDataKit: take a port argument and move off 8000 when it is taken by @danielhanchen in #13072
Load Qwen2.5-Omni checkpoints that name Qwen2_5OmniModel by @danielhanchen in #13087
Fix granite-vision 4-bit generate crash on transformers 4.x (nn.MultiheadAttention out_proj) by @danielhanchen in #13090
Studio: avoid repeated Qwen-Image-2.1 VAE tile OOMs on ROCm by @oobabooga in #13116
fix(studio): repair orphan tool results during chat rendering by @lorenzozanee in #11733
Studio: fall back to the legacy stream when an older backend refuses a new chat field by @danielhanchen in #13130
Studio: use the Vulkan llama.cpp bundle, not CPU, on Windows NVIDIA drivers below CUDA 12.4 by @danielhanchen in #13129
Studio: load GGUF models on Windows when the user profile path is not ASCII by @danielhanchen in #13133
README: install PyTorch before Unsloth on Windows so uv does not resolve an old Unsloth by @danielhanchen in #13134
fix(install): silence the missing AppArmor userns sysctl read by @jayzhou2309 in #13125
Studio: send Decision API images to the MLX engine for Clef by @Lyxot in #13126
MLX decision training: Clef rows and predict take images by @Lyxot in #13127
Studio: keep merged table cells in their columns when reading web pages by @NilayYadav in #13098
Studio: install the llm-compressor-main shadow from a source archive instead of git by @danielhanchen in #13132
Decide accepts_loss_kwargs from the loss head so gradient accumulation gets num_items_in_batch for every model by @danielhanchen in #12946
Keep Dia's audio pad token in fast generate by @danielhanchen in #13128
Studio: speculative decoding for the MLX inference backend by @Lyxot in #13014
Studio: repair ROCm attention packages and bound Qwen-Image-2.1 math prefill by @oobabooga in #13115
Studio: offer vLLM or SGLang when the default engine cannot run a quantized checkpoint by @danielhanchen in #13135
Fix evaluate() under auto padding-free and keep compute_metrics rows per example by @danielhanchen in #13136
GRPO: train on video prompts instead of scoring them as text by @danielhanchen in #13144
Studio: stop with a clear error when a pre-cast text encoder fails and its dense shards were skipped by @KafKafrnZ in #13117
Studio: stop RAG PDF parsing from starving the server on scanned pages by @danielhanchen in #13137
Studio: keep queued prompts when the model is reloaded or switched by @danielhanchen in #13147
Baseline the unsloth-zoo 2026.10.3 and dnspython 2.9.0 findings after review by @danielhanchen in #13143
Studio: keep a failed folder link from spending its grant by @danielhanchen in #13131
Clean machine install: assert the no-git path by what setup.ps1 prints after #12971 by @danielhanchen in #13170
Installer differential: ignore winget's partly filled download bars by @danielhanchen in #13172
Studio: keep annotate marks on PDF pages and slides that remount by @shimmyshimmer in #13119
Fix 4-bit VLM DDP training crash on ranks >= 1 (CheckpointError) by @danielhanchen in #13149
Installers: hash-check the uv fallback script; fail closed on undigested Lemonade assets by @danielhanchen in #13189
Studio: cap every /api write body before auth; keep chat text out of INFO logs by @danielhanchen in #13192
Studio: open http(s): source links without slashes as web links; flag Windows theme downloads by @danielhanchen in #13195
Studio: add a sandbox memory limit setting to Settings > Sandbox by @danielhanchen in #13194
Studio: translate the browser Downloads and video player strings in every locale by @danielhanchen in #13169
Fix the Backend CI test failures after the branch-restore and Omni routing merges by @danielhanchen in #13173
Apply embedding_learning_rate to embeddings under full finetuning by @danielhanchen in #13171
Use int64 row offsets in RoPE and norm kernels by @arcusbuilds in #13121
Hub token identity tests: let the llama-backend stubs take the caller's token after #13005 by @lonexreb in #13022
Studio: stop a linked folder re-indexing failed files forever and let Unlink stop a running sync by @indrajeetapache in #13145
Fix DecisionTrainer crashing on step 1 without a prior inference pass by @danielhanchen in #13176
Studio: add a spell check switch to Settings > General by @KafKafrnZ in #13023
Studio: support Qwen-Image-2.1-Turbo with its 8-step sampling schedule by @oobabooga in #13159
version_compat tests: follow TRL 1.15's chunked_nll alias and GPU-only DPO scorer by @danielhanchen in #13207
Studio/feat: add a run-setting toggle to disable tool-call deduplication by @frostyautumnleaf in #11685
Studio: make OpenBLAS thread limits work on Windows ROCm, and default numpy to up to 8 threads instead of 1 by @danielhanchen in #13048
Unsloth Studio installer (AMD/Linux): use ROCm 6.4 PyTorch on ROCm 6.0 to 6.3 so bitsandbytes training doesn't crash by @TaskPuppyNatani in #10277
Studio: read AGENTS.md into chat instructions by @danielhanchen in #13199
Studio: stop cutting a fetched page down to one article element by @NilayYadav in #13178
Fix case-sensitive Hugging Face projector pairing by @Imagineer99 in #13161
Studio: send each line of a file attached in Chat with files once by @NilayYadav in #13179
Studio: mount sidebar Recents a page at a time so long chat histories stay fast by @oobabooga in #13165
Studio: train tool-calling datasets with their tool list by @NilayYadav in #13180
Studio: turn on agent access (MCP) from Settings > API by @Etherll in #13167
Studio: name the /v1/models VRAM-fit context estimate by @AzizMuminov in #12599
Fix illegal memory access in block-FP8 LoRA kernels on long batches by @danielhanchen in #13141
Studio: show the whole conversation for imported Claude chats by @NilayYadav in #13182
Studio: keep HTML table rows together in knowledge bases and sources by @NilayYadav in #13186
Studio: let small models send an attached image to a mapped MCP tool by @Etherll in #13166
Studio: let @skill mentions work without Code and add bulk skill toggles by @Etherll in #13163
Studio: fill the context bar for Qwen connections by @NilayYadav in #13187
Studio: regenerate a chat's title from its latest turns by @danielhanchen in #13197
Fall back to Unsloth inference for gpt-oss with fast_inference=True by @danielhanchen in #13155
Record the repo id for ModelScope loads so merged_16bit and GGUF export work by @danielhanchen in #13150
Studio: give compare chat the same composer as normal chat by @shimmyshimmer in #13212
Studio: keep the composer expand button in the top corner when a file is attached by @shimmyshimmer in #13198
Studio: cap sd.cpp threads at the cgroup CPU quota and honour UNSLOTH_CPU_THREADS by @danielhanchen in #13213
Studio: treat a Smart App Control block of unsloth.exe like the AppLocker one by @oobabooga in #13168
Studio: list unsloth/Qwen-Image-2.1-Turbo-GGUF in the Images picker by @danielhanchen in #13206
Studio: shorten the vLLM / SGLang install prompt by @danielhanchen in #13216
Studio: fit MCP image mappings in the dialog and keep one field per tool by @Etherll in #13162
Studio: show the installed Lemonade and FastFlowLM versions on the NPU by @oobabooga in #13151
Studio: choose how often update notifications appear, including Unsloth updates by @shimmyshimmer in #13200
Studio: bound Qwen-Image-2.1 attention on Apple Silicon and fail loudly on NaN images by @danielhanchen in #13188
Unsloth Studio (AMD): pick a discrete GPU that fits over a larger APU in automatic GPU selection by @danielhanchen in #13196
install.sh: install XPU PyTorch on Linux when an Intel Arc / Data Center GPU is the only GPU by @danielhanchen in #13193
Studio: name the app Unsloth in the MCP settings copy by @danielhanchen in #13230
Studio (Jetson): keep JetPack's CUDA ahead of pip CUDA in llama-server and backend loader paths by @danielhanchen in #13191
Studio: keep the chat composer typing after a send on WebKitGTK Wayland by @danielhanchen in #13222
Studio: keep tool calls and results when training on ShareGPT tool data by @NilayYadav in #13181
Train text-diffusion LMs: DiffusionTrainer with recipes for DiffusionGemma, LLaDA, LLaDA-MoE, LLaDA2, Dream, Nemotron and SDAR by @danielhanchen in #13157
Fall back to a dynamic KV cache when the static one would not fit by @danielhanchen in #13075
Tests: drop duplicate tests and fix three that checked nothing new by @LeoBorcherding in #13211
Unsloth Studio / Desktop (AMD, Windows): warn at startup when the AMD driver has the RDNA4 idle-eviction bug by @LeoBorcherding in #13208
Studio: keep heading text that sits inside a button on fetched pages by @NilayYadav in #13184
Studio: browser Downloads button matches the toolbar and closes on a second click by @shimmyshimmer in #13226
Studio: keep list numbers in attached Word and HTML files by @NilayYadav in #13185
Studio: pin Qwen-Image 2.1 Turbo directly under Qwen-Image 2.1 in the Images picker by @danielhanchen in #13239
Support TRL 1.14 and 1.15: ORPO/CPO row cap, masked GRPO sampling metrics, cap to 1.15.0 by @danielhanchen in #13203
GRPO tool calling: accept Unsloth chat templates in TRL's response schema and training template lookup by @danielhanchen in #13202
assistant_only_loss with Unsloth chat templates: fall back to response markers by @danielhanchen in #13204
Let TRL AsyncGRPOTrainer train Unsloth models; fix GRPO vision rewards for tool images and TRL 1.15 by @danielhanchen in #13205
Make Unsloth-patched tokenizers and processors picklable by @danielhanchen in #13209
Studio: show context usage for a revisited chat under a cloud model or none by @danielhanchen in #13225
Add opt-in tokenizers 1.0 release candidate encode path by @danielhanchen in #13214
Studio: size a GPU cache for spilled MoE experts with --moe-cache-mib auto by @danielhanchen in #12951
Unsloth Studio / Desktop: export image LoRAs and pick them up again from a custom models folder by @LeoBorcherding in #13146
Chunked mask test: unwrap every layer, including unsloth_zoo's mask wrapper by @danielhanchen in #13248
Studio: keep exponents when the model reads a web page by @NilayYadav in #13183
Chunked mask test: drop the now-unused _swap_function_references import by @danielhanchen in #13257
README: point notebook-only Docker users at the core tag again by @danielhanchen in #13258
Studio: keep the sidebar More flyout shut when a modal opens during its hover delay by @danielhanchen in #13237
Unsloth Studio (AMD, Windows): don't treat a GPU whose multi-arch kernel pack is missing as covered by @danielhanchen in #13231
Tests: fix three checks that fail for reasons other than a regression by @danielhanchen in #13265
Studio: add a tool output length setting to Settings > Chat by @danielhanchen in #13228
Studio installer: drop a flash-attn built for a torch that is gone, and pair torch 2.9 with torchao 0.15.0 by @alkinun in #13247
Disconnect tests: give the hanging client seconds, not 0.2 s, to be reached by @danielhanchen in #13272
OpenClaw 2026.10.1: define the CI agent under keyed agents.entries by @danielhanchen in #13266
chore(deps): apply six validated Studio dependency bumps by @wasimysaid in #13148
Studio: select all and shift-click ranges in the Library by @Padi142 in #13250
Docs: secure-mode Docker comment also binds JupyterLab to 127.0.0.1 by @danielhanchen in #13275
Studio: confine native dataset drops to the account workspace; cap child API key expiry by @danielhanchen in #13273
Studio: title refresh follows the connection that answered the chat by @danielhanchen in #13274
Fix steering follow-ups disappearing when a response is cancelled by @Imagineer99 in #13249
Studio: pick up a GPU whose driver loads after Studio starts by @danielhanchen in #13232
Fix fp16 LoRA training with plain transformers.Trainer (fused CE gradients underflowed) by @danielhanchen in #13152
Packed SDPA rows: one torch varlen_attn call instead of one SDPA call per segment by @danielhanchen in #13221
Studio vLLM: allow Triton non-constexpr globals so DeepSeek-OCR loads by @danielhanchen in #13264
Keep the padding mask for left-padded label-less training forwards by @danielhanchen in #13154
Skip the fused LoRA kernels when a DoRA adapter is active by @vineethsaivs in #13254
Pass GradScaler to GKD chunked JSD with an older-Zoo fallback by @taking-lying-flat in #13190
Treat Docker on WSL2 as WSL when deciding embedding offload by @danielhanchen in #13259
Run the RoPE and NF4 Triton kernels on GPUs older than Volta by @danielhanchen in #13153
Studio: stop charging SWA context checkpoints against VRAM on discrete GPUs by @danielhanchen in #13242
feat(api-monitor): add Load button to API page toolbar by @michupeel in #10378
Studio: check loopback for an existing listener before taking the port by @Nitish-1303 in #11634
[Bug] Images tab only show Reference/Edit, etc sections if its pinned to the sidebar by @aBiR1D in #12767
Studio: name a missing GGUF metadata key instead of blaming the file or memory by @danielhanchen in #13285
Installer: name the missing tools, not Debian packages, on non-apt Linux by @danielhanchen in #13284
Studio: start vLLM on hybrid models whose Mamba state outruns max_num_seqs by @danielhanchen in #13235
Studio: keep auto-load and compare panes off the default engine for compressed-tensors, AWQ and GPTQ checkpoints by @danielhanchen in #13229
Studio (Jetson): size llama-server context from system RAM when nvidia-smi reports [N/A] by @danielhanchen in #13286
Studio: keep a media chat's context usage after a local model switch by @danielhanchen in #13281
Studio: let an MCP call run again after other work changed state by @danielhanchen in #13287
Studio: resolve the FLUX.2-klein 9B base for a local ComfyUI file by @danielhanchen in #13290
Run packed rows on the transformers modeling path through varlen flash attention by @danielhanchen in #13234
Studio: name an MLX or managed-engine token count by the id status publishes by @danielhanchen in #13282
Studio: install and run vLLM on AMD GPUs, on Linux and through WSL on Windows by @oobabooga in #12785
Studio: bound the ChatGPT subscription and NPU disconnect watcher teardown by @danielhanchen in #13293
Studio: move chat with files into a panel above the composer by @shimmyshimmer in #13078
Studio: stop dropdowns jumping on open under Reduce Motion by @shimmyshimmer in #13238
Studio: pair Qwen3.5 / 3.6 vision projectors from another quantizer by @danielhanchen in #13288
Studio: keep the export Save folder input as typed by @jayzhou2309 in #13142
Studio: shared media bar, Open with apps, steady image lightbox by @shimmyshimmer in #13241
Studio: keep export status polling single-flight by @ousamabenyounes in #10679
Studio: draw browser menus, tooltips and toasts over a live native page by @shimmyshimmer in #13122
Run GRPO under FSDP2: no inference mode around FSDP, and an optimizer that steps DTensors by @danielhanchen in #13156
Studio: fix image and video loads on torchao 0.14 (FqnToConfig import error with diffusers 0.41) by @danielhanchen in #13308
Studio installer (macOS): start the Unsloth Studio.app shortcut natively on Apple Silicon by @danielhanchen in #13305
Build the prebuilt CUDA 13 kernel wheels for Windows and ARM (win_amd64, linux_aarch64, win_arm64) by @danielhanchen in #13261
Studio: install prebuilt flash-attn, causal-conv1d and mamba-ssm wheels on Windows and ARM by @danielhanchen in #13262
Studio: torch 2.13 for new Windows x64 cu130 Python 3.13 installs, existing installs keep their torch by @danielhanchen in #13263
Studio (Windows): read a runtime lock held by an elevated Unsloth process as busy, not a broken install by @danielhanchen in #13301
Unsloth Studio: export a trained image LoRA from the Train tab, list exported copies once by @danielhanchen in #13289
Studio: price the int8 ConvRot text encoder on Apple Silicon when the dense one is still cached by @danielhanchen in #13306
Studio: fix torch.compile for image generation on AMD (training config leak, Windows without Visual Studio, max tier) by @danielhanchen in #13227
Studio: reload without MTP when llama-server dies mid-generation on any MTP launch by @danielhanchen in #13316
Studio: keep llama.cpp's Metal fit inside the GPU wired headroom by @danielhanchen in #13319
Intel XPU: 4bit loads without a C compiler by falling back to bitsandbytes' PyTorch quantize by @danielhanchen in #13309
Apply an explicit rope_scaling to the model and to vLLM by @danielhanchen in #13315
Studio: remember Max Tokens on Max as Max, not as the old context length by @danielhanchen in #13223
torchao alias tests: stop leaving a fake diffusers in sys.modules by @danielhanchen in #13336
Studio: pair torch <= 2.8 with torchao 0.15 and gate torchao export on 0.15 by @danielhanchen in #13329
Studio: gate auth.db and rag.db opens and closes against the SQLite 3.51.0-3.51.1 deadlock by @danielhanchen in #13307
Studio: keep loop notes ahead of the MCP image envelope by @ItsRoy69 in #11367
Studio: stop listing Bonsai DSpark drafter files as models by @kevin9327 in #11265
Studio (AMD/ROCm): use MIOpen immediate mode for image and video convs by @danielhanchen in #13323
Studio: explain GGUFs the bundled llama.cpp cannot read, and default Ternary Bonsai to its loadable pack by @danielhanchen in #13304
Keep offloaded embeddings in host RAM when Trainer and accelerate place the model by @danielhanchen in #13320
Studio: list Kimi K3 for Kimi API keys by @danielhanchen in #13300
Studio: let a text Expression column in Data Recipes render empty by @danielhanchen in #13296
Studio: install the Vulkan audio.cpp build on AMD GPUs by @danielhanchen in #13298
Studio: skip the SNAC detokenize probe when the ids are past the vocab by @danielhanchen in #13297
Studio: launch q4_0 instead of iq4_nl KV cache on CUDA/HIP GPUs by @danielhanchen in #13330
Studio: fit an oversized DiffusionGemma GGUF to the GPU in Auto mode by @danielhanchen in #13312
Studio: let YuE2 finish a song instead of cutting it at the asked length by @danielhanchen in #13346
Studio: make Search and Code toggle states clear in dark palettes by @umran666 in #12108
Studio: use a ComfyUI text_encoders file for a local ComfyUI DiT instead of downloading the base repo's encoder by @danielhanchen in #13314
Studio: classify qwen_image21 GGUFs as text-to-image by @abhishekacharya200 in #13333
Studio: Recommended Hub filter and chat picker list FP8 / NVFP4 on capable GPUs, GGUF lists every GGUF by @shimmyshimmer in #13240
AsyncGRPOTrainer with fast_inference: train and generate on one GPU with shared weights by @danielhanchen in #13292
Save LoRA-GA adapters relative to the original base so they reload correctly by @danielhanchen in #13322
Studio: keep prompt cache on seeded tool-loop continuations by @indrajeetapache in #10707
Studio: load and run DeepSeek-OCR and DeepSeek-OCR-2 by @danielhanchen in #13310
fix(studio): expose tool settings for external models by @Hal9999 in #13251
Studio: stop offscreen reasoning transcripts listening to thread scroll by @danielhanchen in #13046
Studio: continue LoRA training from a saved adapter checkpoint by @indrajeetapache in #13158
Keep token fields when DiffusionTrainer converts a config by @vineethsaivs in #13335
Studio: stop a chat switch mid-reply crashing the app on a stale message index by @danielhanchen in #13352
Mask packed document boundary labels for TRL older than 0.23.1 by @danielhanchen in #13324
Fall back instead of failing import when torchcodec cannot load its native library by @danielhanchen in #13348
Keep modules_to_save trainable when reloading a LoRA checkpoint by @danielhanchen in #13303
Studio: lighter chat with files border in light mode by @shimmyshimmer in #13362
Unsloth Desktop: redact dashed sk- keys, JWTs and JSON credentials in support diagnostics by @LeoBorcherding in #13260
Studio: bound a live Thinking block to its newest lines by @danielhanchen in #13086
Studio: show the compaction count in the context meter by @danielhanchen in #13355
Unsloth Studio / Desktop: colour a companion download's bar by model, text encoder and VAE by @LeoBorcherding in #13112
Studio: make the web UI installable as a PWA by @danielhanchen in #13360
Fix Q-GaLore checkpoint weights and stacked MoE grouping by @taking-lying-flat in #13276
Never train with a KV cache, so padding-free rows keep their document boundaries by @danielhanchen in #13177
Studio: keep a wrapper-less Gemma value a string unless JSON reads it as a number by @MohammadHijjawi97 in #13279
Train embed_tokens and lm_head in FastModel after add_new_tokens by @danielhanchen in #13057
Studio: run React and TypeScript code blocks as live previews by @Etherll in #13039
Studio: stop refusing C:\path\app.exe as the blocked command '.' by @danielhanchen in #13347
AsyncGRPO vLLM server: raise the listen backlog past socketserver's 5 by @danielhanchen in #13381
Studio: upload files straight into a chat's folder by @danielhanchen in #13366
studiobench: carry on with image_upload when the click opened the menu but timed out by @danielhanchen in #13377
Studio: Deep Research Retry continues with the current time budget by @danielhanchen in #13358
Studio Hub: filter models by parameter count, context length and publisher by @danielhanchen in #13363
Keep a user-set UNSLOTH_RETURN_LOGITS=1 across trainer.train() by @danielhanchen in #13311
Studio: let LAN access bind only the addresses you choose by @KafKafrnZ in #12839
Studio: per-account model allowlist for multi-user installs by @danielhanchen in #13359
Studio: optional detailed Live monitor with GPU load, temperature and power by @danielhanchen in #13351
Studio: add an embedding device setting for RAG by @danielhanchen in #13365
Studio: torch 2.13 for new Linux aarch64 cu130 Python 3.13 installs, existing installs keep their torch by @danielhanchen in #13380
Studio: forward llama-server /metrics, /slots, /tokenize and /detokenize by @danielhanchen in #13354
Studio: add an opt-in setting to read replies aloud automatically by @danielhanchen in #13357
Forward output head bias in GRPO policy scoring by @taking-lying-flat in #13302
Revert "Studio: keep the composer expand button in the top corner when a file is attached" by @shimmyshimmer in #13394
Studio: classify the 10 new audio.cpp-gguf folders as needing a newer runtime by @danielhanchen in #13393
Studio: screen a quoted program path that cmd /c or start launches by @danielhanchen in #13386
Studio: Working... placeholder with a soft glint by @shimmyshimmer in #13382
Tests: keep five tests off what the host has installed or downloaded by @danielhanchen in #13395
Honor padding masks in Llama and Qwen3 cached attention by @Arthur031221 in #13256
Studio: hide the response-details badge without a fade by @shimmyshimmer in #13397
Switch between training and inference mode in one tree walk by @danielhanchen in #13374
Studio: let a project run its tools in an existing local folder by @danielhanchen in #13353
Perf/scroll performance by @Sneakr in #13255
Studio: let a python/terminal call ask for a lower output cap by @csarushan1729 in #10398
feat(dataprep): add audit_supervision, a pre-flight report of what an SFT dataset actually supervises by @RubenHaisma in #10852
Keep packed documents apart on Unsloth's own models (flash path, eval, cached eval) and match transformers' Mistral window by @danielhanchen in #13313
Sandbox delete test: only the chat's detached tree counts as the delete by @danielhanchen in #13403
Studio: sd.cpp on Macs, CPU backend on a paravirtual GPU and no direct conv on Metal by @danielhanchen in #13369
Studio: refuse LTX-2.5 instead of loading it as LTX-2 by @danielhanchen in #13370
Studio MPS: opt-in fp16 compute with fixed headroom for Z-Image by @danielhanchen in #13375
Studio: judge hosted pre-quant checkpoints before downloading them, and make the int8 fingerprint torchao-version-neutral by @danielhanchen in #13339
Studio: refresh the context meter when a tool loop hands off to a tool by @danielhanchen in #13356
Studio: opt-in MLX image engine for Apple Silicon (Z-Image) by @danielhanchen in #13379
Allow offload_layers = "auto" under a DDP launch by @danielhanchen in #13364
Opt-in DDP batched_grad_copy (UNSLOTH_DDP_BATCHED_GRAD_COPY=1) by @danielhanchen in #13350
Studio: document the intentional torch and torchcodec floor gap by @RudraBhaskar9439 in #10635
Studio: Wan2.1 T2V 1.3B and 14B video families by @danielhanchen in #13376
Packed rows on sliding-window layers take varlen on the transformers path, plus opt-in FA4 / FA3 for packed training by @danielhanchen in #13326
FastLanguageModel.for_training: undo the UNSLOTH_RETURN_LOGITS=1 that generate() forces by @danielhanchen in #13392

New Contributors

@SulimanAbdulrazzaq made their first contribution in #12011
@frostyautumnleaf made their first contribution in #11685
@michupeel made their first contribution in #10378
@Nitish-1303 made their first contribution in #11634
@aBiR1D made their first contribution in #12767
@kevin9327 made their first contribution in #11265
@abhishekacharya200 made their first contribution in #13333
@Hal9999 made their first contribution in #13251
@Arthur031221 made their first contribution in #13256
@csarushan1729 made their first contribution in #10398
@RubenHaisma made their first contribution in #10852

Full Changelog: v0.1.905-beta...v0.1.906-beta

unslothai/unsloth v0.1.906-beta Browser Use + Qwen-Image-2.1-Turbo on GitHub