unslothai/unsloth v0.1.908-beta on GitHub

Run NVIDIA's new Nemotron 3.5 models and Meta's new Muse Spark Open locally in Unsloth. New: text-diffusion LLMs, DeepSpeed training and better prompt caching for coding agents.

Highlights

NVIDIA Nemotron 3.5 Bolt: Run on 24 GB of memory. Guide
NVIDIA Nemotron 3.5 Super VL: 120B (12B active) Guide
Meta Muse Spark Open: 816B (42.7B active) - runs on 256GB Guide
Better prompt caching for coding agents: Claude Code, Codex etc. on local GGUF models re-process far less of each prompt.
Browser use: Always allow properly works, cursor, and separate tabs for each chat.
Images through your connections: generate with sd-server, Ollama, Lemonade and other image servers.
Diffusion LLMs: train and chat with LLaDA, Dream, SDAR and more, with a live denoising canvas.

Nemotron 3.5

Run NVIDIA Nemotron 3.5 Super VL, a 120B model with 12B active parameters and up to 1M tokens of context, built for document understanding, computer use and long-running agents. Super VL runs in about 64 GB of memory at 1-bit, or 128 GB at 4-bit. Guide · GGUF
Run NVIDIA Nemotron 3.5 Bolt, a 74B model with 9B active parameters and 128K context, compressed from Super VL. Bolt runs in about 32 GB of memory at 1-bit or 2-bit. Guide · GGUF
Nemotron 3.5 Super VL generates about 5x faster than Inkling Small at 100K input with similar intelligence.
By default, Unsloth runs Nemotron 3.5 Bolt with a DSpark drafter (about 1 to 2 GB more memory) and Super VL with its built-in MTP for faster output. Both can be turned off in settings. MTP Guide

Muse Spark Open

Run Meta's Muse Spark Open, an 816B MoE model with about 42.7B active parameters and 1M tokens of context, built for multimodal reasoning, coding and agentic workflows. Muse Spark Open runs in about 256 GB of memory at 2-bit. Guide · GGUF
Muse Spark Open takes text, images, video and audio as input, with reasoning effort from low to xhigh.
Muse Spark Open was trained on long-horizon coding tasks, including whole-repository generation, large end-to-end projects and auto-research.

Training

Train and chat with text-diffusion LLMs in the Unsloth app, including LLaDA, Dream, SDAR, LLaDA2, Nemotron-Labs-Diffusion and DiffusionGemma. Chat replies show a live denoising canvas.
Train with DeepSpeed ZeRO-2 and ZeRO-3, including 4-bit QLoRA, full fine-tuning, DPO and CPU offload. LoRA training with DDP and gradient checkpointing also no longer crashes.
GRPO with fast_inference now runs on Nemotron 3 Nano and block-FP8 MoE models, sharing weights with vLLM.
On the Train page, stop early when eval loss stops improving and train on pre-tokenized datasets. Random Seed now also sets the LoRA starting weights.

Performance

When Claude Code, Codex or OpenCode runs on a local GGUF model, follow-up turns re-process much less of the conversation, and a new Claude Code session's first reply starts up to 4x sooner.
In Python, single-prompt model.generate is up to 2.7x faster at 32K context for Llama and Qwen3 models.
In Python, generating 2 to 4 prompts at once with a bitsandbytes 4-bit model is up to 2.8x faster on some NVIDIA GPUs.
GRPO rollouts with fast_inference on Nemotron 3 Nano are up to 30% faster by default.

Browser use

See a cursor where the model clicks and types, or turn it off in Settings > Browser > Browser use > Show cursor.
Browser use prompts now have Always allow, which approves that kind of action on a site for the rest of the chat. Scripts still ask every time.
Each chat keeps its own browser tabs, and in Desktop, browser use keeps working in chats that are not on screen.
In the web UI, browser pages remember site preferences between pages, load data from their own site, and play videos full screen.

Images + video

Generate images on the Images page through a connected image server, such as sd-server, Ollama, Lemonade, vLLM, llama.cpp or a custom OpenAI-compatible URL.
Image and video models now skip text-encoder layers they never use, saving memory on every load.

Chat + Desktop

Choose where Unsloth installs on Windows from Desktop's first-run screen.
Re-embed your documents after switching the embedding model, without uploading them again.
Chat with files can collapse to a row of pills, its citations open a PDF at the cited page, and newly added Word and HTML files keep their list numbers.
Compact a conversation from the chat header menu, and Rewind no longer asks for confirmation when no files changed.

Models + settings

When Claude Code, OpenCode or Codex runs on a cloud model, add --auto-route to unsloth start to send routine coding work, such as reading files, searching and running commands, to your local model.
Codex can use its MCP and sub-agent tools with local models, and Claude Code, Codex and OpenCode can run on a model from Settings > Connections with unsloth start --model.
Name a model once in Settings > API > Model aliases, so apps using the Unsloth API keep working when the model behind the alias changes.
vLLM connections and Unsloth's own vLLM engine support thinking effort levels and Preserve thinking, and the engine now auto-compacts long chats. vLLM and SGLang loads can also set an FP8 KV cache, a GPU memory limit and a seed.
Pick 2-bit, 4-bit, 6-bit or 8-bit from MLX repos that publish several precisions, the way you pick GGUF quants. Xing4.0 and MiniCPM-V 4.7 now run and train with MLX.

Hardware compatibility

Windows PCs with NVIDIA drivers older than CUDA 12 install the Vulkan or CPU llama.cpp build instead of failing.
AMD GPUs that ROCm does not support, such as RDNA 1 and Polaris, now get the Vulkan llama.cpp build even when ROCm is installed or another AMD GPU supports it.
GGUF image generation runs on Vulkan for AMD and Intel Arc GPUs when PyTorch is CPU-only.
Windows installs on PyTorch 2.10 and 2.11 get prebuilt flash-attn, causal-conv1d and mamba-ssm.
Multi-GPU machines that mix GPU generations now pick the attention backend for each GPU, and Hopper GPUs with conda-forge xformers no longer crash.

Download Unsloth Desktop

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

Platform	Link
Windows	Download
macOS	Download
Linux x64 / Ubuntu (deb)	Download
Linux ARM64 / Ubuntu 24.04+ (deb)	Download
Linux x64 (AppImage)	Download
Windows ARM64	Download

What's Changed

fix: update Codex client version to restore model discovery by @Imagineer99 in #13581
Studio: changing the thinking setting also resets Expand thinking by @shimmyshimmer in #13625
Bump install.sh / install.ps1 pins to unsloth>=2026.10.5 by @danielhanchen in #13627
Studio MCP: never run server-side tools from the chat tool by @danielhanchen in #13605
Nemotron-H fast_inference via vLLM weight sharing (fallback on older unsloth_zoo); drop vLLM-only kwargs on fallback by @danielhanchen in #13440
Diffusion LM training: compiled chunked loss with 0 graph breaks, plus per-family samplers by @danielhanchen in #13341
Studio: offer the precision folders of an MLX repo like GGUF quants by @danielhanchen in #13565
Studio: keep ticked Word form check boxes in Chat with files by @NilayYadav in #13485
tests: click the inline Delete button #13615 put back on the action bar by @danielhanchen in #13634
Studio: hold the forced aligner against a folder delete; match compound audio.cpp size scopes in order by @danielhanchen in #13633
Studio: vLLM thinking effort levels, preserve thinking and auto-compact by @danielhanchen in #13571
Studio: stop dropping the page text after a menu with no closing tag by @NilayYadav in #13489
Re-measure the Studio startup budget after the audio selector and thinking merges by @danielhanchen in #13640
Studio: let the Python tool open a SQLite file by @NilayYadav in #13487
FP8 LoRA: dequantize down_proj in the backward like gate and up by @danielhanchen in #13270
FP8 on ROCm: use the dequant path on archs without e4m3fn Triton support by @danielhanchen in #13331
Freeze the setup-time heap during Trainer.train(); add resume parity and memory-flat GPU tests by @danielhanchen in #13332
Studio: train and chat with every text-diffusion LM family by @danielhanchen in #13342
Studio: keep the local Claude agent from stopping on a small context window by @NilayYadav in #13477
Studio: run Laya on the GPU by default and free it on unload by @NilayYadav in #13479
Studio: pass Codex MCP and sub-agent tools to the model on /v1/responses by @NilayYadav in #13483
DDP LoRA: default find_unused_parameters=False so gradient checkpointing does not crash by @danielhanchen in #13267
Studio: KV cache dtype, GPU memory limit and seed for vLLM and SGLang loads by @danielhanchen in #13596
Diffusion LMs: fast inference, live denoising frames and in-loop cancellation by @danielhanchen in #13592
fast_cross_entropy_loss: all-masked labels give loss 0 instead of NaN by @danielhanchen in #13325
Studio: reload a Clef as soon as the training run that unloaded it ends by @NilayYadav in #13482
Studio: send routine coding work to the local model with unsloth start --auto-route by @NilayYadav in #13478
tests: accept a positional encoding on Path.read_text / write_text by @danielhanchen in #13646
QAT: fall back to the module forward for torchao NVFP4 / MX linears in fused LoRA by @danielhanchen in #13280
Tests: keep the zoo num_proc anchor tests meaningful once the zoo adopts the edit by @danielhanchen in #13269
Studio: point unsloth start at the right model when several are loaded by @NilayYadav in #13480
Studio: live denoising canvas for transformers diffusion LMs by @danielhanchen in #13593
FP8 MoE + fast_inference: refuse unsupported checkpoints before vLLM loads, with a hint that works by @danielhanchen in #13277
Save each unsloth train run in its own folder by @NilayYadav in #13484
Studio: keep data recipe seed inspect inside a managed account's workspace by @danielhanchen in #13611
GRPO DDP gradient sync: copy each averaged bucket back with one foreach_copy by @danielhanchen in #13268
Native decode: read GQA KV heads in place instead of copying them per query head by @danielhanchen in #13271
Studio: key the i2v availability cache on the diffusers module, not its id by @danielhanchen in #13667
Studio CI: click the composer itself in the dismiss_on_composer probe by @danielhanchen in #13670
Studio UI test: wait for the chat shell to settle, not just the composer by @danielhanchen in #13672
End the last gpt-oss training turn with <|return|>, not <|end|> by @NilayYadav in #13657
Studio: LTX-2.5 defaults follow ComfyUI (hosted INT8-ConvRot DiT, opt-in int8 text encoder) by @danielhanchen in #13618
Studio: Wan2.2-TI2V-5B LoRA training by @danielhanchen in #13654
Studio: declare the Hub's /api/engines retry in the log budget by @danielhanchen in #13682
Studio: Collapse thinking's chevron points up by @shimmyshimmer in #13669
Studio: size the composer to the message column at every width and zoom, and stop clipping thinking text by @shimmyshimmer in #13675
Studio: make the first-token waiting limit configurable by @Imagineer99 in #13390
Model-config UI test: treat a /validate that never reports its end as a lost event by @danielhanchen in #13689
Studio: Windows torch 2.11 kernels from our prebuilt wheels by @danielhanchen in #13564
Studio: Chat with files search row lists its documents, which open in the attachment viewer by @shimmyshimmer in #13681
Studio: one scrolling attachment row in a chat, even spacing above the input by @shimmyshimmer in #13678
Studio: train structured recipe columns as JSON, not as Python dicts by @NilayYadav in #13635
Studio: Skills pill stays with every skill unticked; its X or the + menu turns Skills off by @shimmyshimmer in #13687
Keep torch's bundled flash on Blackwell for short documents under UNSLOTH_FLASH_ATTENTION_IMPL=auto by @danielhanchen in #13631
Declared-polls probe: count only requests this Studio serves by @danielhanchen in #13690
Studio: collapse Chat with files to a row of pills, with Show all by @shimmyshimmer in #13686
Studio: drop the stray divider above a tool's created-files row by @shimmyshimmer in #13692
Studio: simpler Rewind confirm, skipped for conversation-only rewinds by @shimmyshimmer in #13676
Studio: removing a composer attachment no longer reloads the others by @shimmyshimmer in #13677
Studio: text #0f1114 in light mode and #fcfcfc in dark, thinking text easier to read by @shimmyshimmer in #13693
Studio: chat with files citation chips open a PDF at the cited page by @shimmyshimmer in #13691
Studio: let read aloud go past 2048 tokens and warn when cut short by @NilayYadav in #13648
Studio: keep SSE MCP servers as SSE when importing a data recipe by @NilayYadav in #13659
Studio: reorder the chat + menu, with Deep research and Browser use by @shimmyshimmer in #13679
Studio: let JSON-reply API requests still call their tools on MLX models by @NilayYadav in #13658
Studio: Settings dropdowns glow in the pane colour in dark mode by @shimmyshimmer in #13703
Pack Qwen3.5 text SFT with the processor, and fix eval_loss on packed text runs by @danielhanchen in #13624
Fix main CI: rewind and Skills strings, token count harness, deep research contract, zoo 2026.10.5 scan baseline by @danielhanchen in #13701
CI: coalesce main-push runs per workflow so merge bursts stop outrunning the runner pool by @danielhanchen in #13705
Studio: add early stopping patience on eval loss by @danielhanchen in #13626
CI: bound CodeQL overlay-base cache churn, unstarve the cache janitor by @danielhanchen in #13704
Unsloth Studio: redact the credential after an Authorization scheme in research queries by @LeoBorcherding in #13499
Fix DPO on models split across GPUs with TRL 1.15's fused LM head by @danielhanchen in #13641
Keep the pad embedding row when the tokenizer emits the pad token by @danielhanchen in #13628
Fix Voxtral processor saving and Voxtral Realtime training by @YashwanthNomula in #12647
Skip xformers FA3 on H100 when the build has no sm_90a kernels by @shimmyshimmer in #13619
Skip the per-layer RoPE stream sync unless the model spans several devices by @danielhanchen in #13609
Train with DeepSpeed ZeRO-2 and ZeRO-3 by @danielhanchen in #13587
CI: cancel a pull request's leftover queued and running runs when it closes by @danielhanchen in #13709
CI: overlap browser installs and cache restores with Install Unsloth in the UI smoke jobs by @danielhanchen in #13707
CI: run lint checks in parallel and overlap independent setup in Windows smoke and Backend CI by @danielhanchen in #13710
Chat UI test: close the stale page browser-side instead of waiting on its renderer by @danielhanchen in #13708
Studio: use browser Always allow rules during composer Full access by @wasimysaid in #13700
Studio Desktop: keep a maximized window above the Windows taskbar by @danielhanchen in #13647
CI: let staging repos save dependency caches on upstream-main by @danielhanchen in #13711
CI: wait on events instead of fixed sleeps in Playwright smokes, backend tests and Stop Unsloth steps by @danielhanchen in #13712
fix(studio): anchor relative PATH entries during setup by @sahacha in #13606
Studio frontend: raise the startup bundle budget by the measured growth since the last raise by @danielhanchen in #13718
Studio: keep the prompt cache for API clients and coding agents by @danielhanchen in #13645
Studio: keep the browser page background when widening the panel by @shimmyshimmer in #13723
CI: run the 34 mocked GPU-selection tests, drop dead zoo deselects, print skip reasons by @danielhanchen in #13706
Studio Hub: Newest and Recently updated list more than Unsloth models under Recommended by @danielhanchen in #13632
Move the Unsloth Studio MLX pins to mlx-vlm 0.7.8 by @Lyxot in #13699
feat: add compact conversation to the chat menu by @Lyxot in #13579
Studio: keep the composer draft across chats, pages and tabs by @shimmyshimmer in #13722
Lint CI: scan committed files, not the tree the background lanes write into by @danielhanchen in #13728
Studio: stop Expand thinking from restyling the whole trace by @oobabooga in #13665
Studio: show the hovered icon's tooltip when moving between adjacent icons by @danielhanchen in #13735
Prebuilt CUDA wheels: torch 2.10 for Windows x64 by @danielhanchen in #13740
Studio installer (AMD, Linux): Vulkan llama.cpp for RDNA 1 and Polaris when ROCm is installed by @danielhanchen in #13694
Studio (macOS): read the Metal working set from torch when MLX is missing by @danielhanchen in #13695
Lint CI: skip .lanes in the whole-tree scans by @danielhanchen in #13736
Revert "Studio Hub: Newest and Recently updated list more than Unsloth models under Recommended" by @shimmyshimmer in #13744
Studio: search and load large MCP tool catalogs on demand by @danielhanchen in #13601
Studio: default export load to offload_layers=auto by @indrajeetapache in #13720
MLX: Clef Omni in Unsloth Studio, with audio in the Decision API and decision training by @Lyxot in #13538
Composer browser simulations: log the browser process on the retry by @danielhanchen in #13745
Hybrid packing: keep the varlen forward working after accelerate's unwrap_model (Qwen3.5 / 3.6 GRPO) by @danielhanchen in #13721
Pick the attention backend per device on mixed-generation multi-GPU hosts by @danielhanchen in #13698
Studio: run native image generation on Vulkan when torch is CPU-only on an AMD or Arc GPU by @danielhanchen in #13697
Nudge-retry cancel test: bound the hang, not the runner's speed by @danielhanchen in #13750
Compute RMSNorm weight gradient when the norm weight is trainable by @danielhanchen in #13594
Studio: send Gemini the image just attached in a chat full of images by @NilayYadav in #13636
Studio: prevent tooltips from hidden run settings by @wasimysaid in #13614
Studio: say why Sandbox High is unavailable when there is nothing to install by @danielhanchen in #13732
Decode small 4bit batches with the NF4 GEMV instead of dequantizing every token by @danielhanchen in #13642
Studio: keep a Claude session working after an image on text-only GGUFs by @NilayYadav in #13652
Studio: search for llama.cpp and other file names in the address bar by @NilayYadav in #13661
Studio: open the Images and Audio workflows from the collapsed sidebar rail by @danielhanchen in #13731
Keep the residual in Gemma and Gemma2 float32 fast decode by @vineethsaivs in #13643
Keep the Docker notebooks and the models they save after docker rm by @NilayYadav in #13660
Studio: keep Start Export off for a Hub push until the repo is named by @NilayYadav in #13651
Studio: keep mid-conversation Anthropic system messages in place so new sessions reuse the prompt cache by @danielhanchen in #13727
Studio: add an Audio API card to Settings and open it from Audio pages by @Etherll in #12821
Studio: move the RAG embedder to CPU when its GPU llama-server keeps crashing by @danielhanchen in #13739
Studio: forget deleted audio models once the Hub cache listing updates by @Etherll in #13630
Studio: report prompt cache hits on the Messages and Responses APIs, and ask vLLM and SGLang for them by @danielhanchen in #13741
Fix installer warning colors on light backgrounds by @Imagineer99 in #13572
Studio: keep a threadless conversation's cut between requests so the prompt cache is reused by @danielhanchen in #13743
Studio: give a killed llama-server time to exit before a reload fails by @danielhanchen in #13622
Studio: let browser panel pages call their own site's API by @oobabooga in #13664
Fix Gemma2 sliding-window boundaries across attention backends by @taking-lying-flat in #13730
Studio: ask before a desktop Update stops downloads in progress by @NilayYadav in #13650
Studio frontend: raise the startup bundle budget by the measured growth since the last raise by @danielhanchen in #13752
Studio: seed the LoRA init with the Random Seed setting by @NilayYadav in #13637
Studio: keep the start of an attached file larger than the context window by @danielhanchen in #13620
Studio: render numbers wrapped in dollar signs as math, not prices by @NilayYadav in #13649
Studio: deleting a GGUF row no longer deletes other builds at the same quant by @alkinun in #13368
Studio: re-embed documents after switching the embedding model by @Lyxot in #13702
Format studio/backend/routes/inference.py back to the hook's fixed point by @danielhanchen in #13765
fix(studio): preserve chat tool selections across model unloads by @Sneakr in #13621
Studio: install the newest audio.cpp release and offer it in updates by @Etherll in #13589
Studio passthrough: read the backend's identity through getattr in the cut scope by @danielhanchen in #13769
Studio: raise the startup bundle budget by #13589's measured growth by @danielhanchen in #13789
Studio: keep the desktop browser page live when it opens as a dialog by @danielhanchen in #13734
Studio: cache Edit-mode text encoding by reference image content by @kevin9327 in #13683
Studio: Windows torch 2.10 kernels from our prebuilt wheels by @danielhanchen in #13742
Measure the resume glyph's centring in viewBox units, not painted rects by @danielhanchen in #13799
Studio: name the Windows update that adds MXC's built-in container by @danielhanchen in #13798
Prebuilt CUDA wheels: torch 2.12 for Windows x64 by @danielhanchen in #13807
Give the data settings driver's cold first load the same 120s as its mount wait by @danielhanchen in #13809
Studio: stopping a model load leaves no Starting model status behind by @shimmyshimmer in #13803
Studio: unload a model under the id the API lists for it by @danielhanchen in #13810
Unsloth Studio / Desktop (AMD, Windows): install the Vulkan llama.cpp prebuilt when device 0 has no HIP bundle by @LeoBorcherding in #13680
Pin vermin: 1.9.0 fails the backend floor lint on valid 3.5 syntax by @danielhanchen in #13818
Studio: let YuE2 songs run to their own ending instead of 1.5x the length by @danielhanchen in #13371
unsloth start opencode: never pass --model on a bare launch; OpenCode 2 rejects it by @YogeshSeeni in #13783
Studio: read the code behind a GitHub file link in web search by @NilayYadav in #13754
Studio: each chat keeps its own browser tabs, and the browser agent works on in chats off screen by @danielhanchen in #13800
Studio: keep a Connection's thinking between tool calls in the same answer by @oobabooga in #13785
Studio: install the Vulkan or CPU llama.cpp bundle on Windows NVIDIA drivers older than CUDA 12 by @oobabooga in #13671
Studio: tighter hover pill on the chat model selector by @shimmyshimmer in #13817
Studio: let videos in the browser panel go full screen by @NilayYadav in #13759
Studio: keep transcribing past a damaged audio frame by @NilayYadav in #13757
Score a 16-bit fine-tune with unsloth eval in 16-bit, not in 4-bit by @NilayYadav in #13758
Studio: tell the model when a site's bot protection blocks a page fetch by @NilayYadav in #13760
Studio: train ShareGPT tool calls natively on processor-loaded models by @danielhanchen in #13824
Studio: use native Linux browser tabs with GTK overlay sizing by @wasimysaid in #13717
Studio: add Always allow to browser agent prompts by @NilayYadav in #13777
Studio: let tool results use the window when auto-compaction is off by @danielhanchen in #13811
Studio: keep list numbers in Word and HTML knowledge base files by @NilayYadav in #13755
Studio: show the browser agent cursor where it clicks and types, with a Settings toggle by @shimmyshimmer in #13729
Studio: finite MPS memory cap and sd.cpp --mmap on macOS by @danielhanchen in #13763
Studio Hub: drop the compact list view, make grid and split views responsive by @shimmyshimmer in #13822
Studio: let the Linux tool sandbox use bubblewrap on NixOS and in FHS launchers by @oobabooga in #13768
Return weight and bias gradients from the fast LayerNorm, and support bias=False by @danielhanchen in #13801
Studio: probe the regional diffusion compile on ROCm instead of refusing it by @danielhanchen in #13774
Studio: stop re-prompting a finished tool answer that labels its results by step by @oobabooga in #13653
Studio: generic text-encoder dead-tail truncation for diffusion and video by @danielhanchen in #13767
Studio: keep sd.cpp direct conv off on Vulkan and ROCm by @danielhanchen in #13773
Studio NVFP4: one-pass activation quantiser for torchao's to_nvfp4 by @danielhanchen in #13804
Studio diffusion: learned activation peaks, OOM retry in group offload setup, pinned GGUF host copies, tied-weight groups by @danielhanchen in #13724
Studio: add an opt-in Settings switch to hide models that do not fit by @danielhanchen in #13814
Studio: say when a browser panel page needs the system browser by @NilayYadav in #13761
Read the Tauri command list inside the gated invoke_handler closure by @danielhanchen in #13836
Studio: keep site settings across pages in the browser panel by @NilayYadav in #13762
Studio: start audio.cpp options at each model's recommended values by @Etherll in #13764
Allow norm layers in FastLanguageModel modules_to_save by @danielhanchen in #13805
Studio: generate images through an external connection by @NilayYadav in #13756
Studio: Hub Run refuses audio models that neither Audio nor Chat can run by @danielhanchen in #13812
Studio: accept Parquet files as local Data Recipe seeds by @danielhanchen in #13839
Studio Desktop: choose where Unsloth installs on Windows by @danielhanchen in #13446
Studio Hub: one filter popover for capabilities, size, context and uploader, composer-style menus by @shimmyshimmer in #13749
Trim comments in tests by @danielhanchen in #13844
AsyncGRPOTrainer: optional SAO-style importance sampling mask by @danielhanchen in #13820
Studio: stop llama-server crashing on macOS on Ctrl+C by @NilayYadav in #13782
Studio: train on pre-tokenized datasets (input_ids / attention_mask / labels) by @danielhanchen in #13816
Studio: keep the chat template of Mac LoRAs trained on multimodal base models by @danielhanchen in #13826
Fix bfd_split truncating overflow before packing by @taking-lying-flat in #13770
Studio: model aliases for the OpenAI-compatible API by @danielhanchen in #13361
Studio: int8 weight-only transformer quant on Apple Silicon through the Metal int8 kernel by @danielhanchen in #13344
Studio: stop listing a GGUF's borrowed text encoder and VAE as a model in the picker by @LeoBorcherding in #13778
Studio: run coding agents on a Connection model by @NilayYadav in #13481
Studio recipes: keep a finished run's pump off the next run by @danielhanchen in #13845
unsloth run: add --engine-parallelism and --gpus for multi-GPU serving by @danielhanchen in #13838
Run Linux frontend browser checks in two parallel groups by @heychaeyin in #13813
Studio: send audio.cpp request fields in the form each model reads by @Etherll in #13791
Fix DiffusionGemma sampled-prefix cache retention by @taking-lying-flat in #13780
Studio: accept sk-unsloth API keys in the x-api-key header by @danielhanchen in #13842
Studio: let Orpheus TTS pick one of its built-in voices by @danielhanchen in #13853
Pass layers_pattern from FastLanguageModel.get_peft_model to LoraConfig by @vineethsaivs in #13751
Load the zoo Hub download helpers without unsloth_zoo's package init by @danielhanchen in #13849
Fix SDAR empty-mask resampling to match the reference by @taking-lying-flat in #13771
Add dtype option to save_pretrained_merged for float32 merges by @danielhanchen in #13823
fast_inference: Gemma 4 Unified, Muse Glimmer, dense Gemma 4; 4-bit fallback when no -bnb-4bit repo by @danielhanchen in #13825
Keep fp32-only tensors (MoE router bias) in fp32 on 16-bit saves of non-PEFT models by @danielhanchen in #13819
MLX stub: answer zoo's fused_verify_attention by @danielhanchen in #13863
Studio Desktop: open the OS folder dialog to add a model folder by @danielhanchen in #13843
Studio: serve /v1/completions from MLX models by @Lyxot in #13796
Prebuilt CUDA wheels: compress the fatbins (flash-attn 1.49 GB -> ~285 MB) by @danielhanchen in #13854
Kernels: request enable_fp_fusion=True explicitly, ahead of Triton 3.9's False default by @danielhanchen in #13828
Desktop: explain Windows Long Paths before its UAC prompt by @danielhanchen in #13523
Studio: install causal-conv1d and mamba-ssm only when a model needs them by @danielhanchen in #13861
Add UnlimitedOcrTextModel to the fused-MoE conversion snapshot by @danielhanchen in #13867
Fine-tune and export native MTP heads and DSpark drafters by @danielhanchen in #13719
fix(studio): show a download start the moment it is requested, and ke… by @Sneakr in #13738
fast_inference: switch Gemma 4 to Unsloth inference on GPUs below sm80 by @danielhanchen in #13864
Studio: info stamp on image downloads, recipe in every format, load settings from an image by @danielhanchen in #13444
Compile each transformer block as one region for every FastModel model (LoRA and QLoRA) by @TarzanZhao in #10744
Format tests/test_triton_fp_fusion.py back to the hook's fixed point by @danielhanchen in #13887
Studio: stream /v1/responses from MLX and other in-process models by @Lyxot in #13794
Studio Hub: Recommended lists safetensors repos only past 30 likes, always open on Discover by @shimmyshimmer in #13886
Docker ROCm Studio image: run GGUF chat on the GPU through Vulkan by @danielhanchen in #13852
NF4 GEMV: drop K masks on even K and gather the NF4 table from registers by @danielhanchen in #13829
FP8 rowwise GEMV: int64 row offsets past 2^31 elements, and skip it on Blackwell with Triton 3.3 by @danielhanchen in #13830
Stream the chunked cross entropy forward logits with evict_first by @danielhanchen in #13893
Studio: serve /v1/messages and count_tokens from MLX and other in-process models by @Lyxot in #13795
MoE grouped GEMM: tl.range for the dX expert loop, drop the dead dW FLATTEN by @danielhanchen in #13831
Launch eager RMSNorm and LayerNorm kernels through triton_launch by @danielhanchen in #13889
Studio Hub: show the Meta logo on every Unsloth Muse re-upload by @shimmyshimmer in #13896
Fix FastLanguageModel fused CE loss under FSDP2 with an untied lm_head by @danielhanchen in #13827
Studio: Windows torch 2.12 kernels from our prebuilt wheels by @danielhanchen in #13808
Studio: train a speculative decoding drafter with fine-tunes and export it for llama.cpp and vLLM by @danielhanchen in #13806
Frontend CI: keep grouped browser logs through a second cancel signal by @danielhanchen in #13894
Studio: support NVIDIA Nemotron 3.5 Bolt and Super GGUFs by @danielhanchen in #13899
MoE LoRA: keep native per-expert Linears for an explicit target_modules list on transformers 4.x by @danielhanchen in #13851
Studio: let Run settings load a Hugging Face model in 16-bit by @danielhanchen in #13841
Studio: support Muse Spark Open GGUFs by @danielhanchen in #13903
Fix main's red Backend CI, repo and picker contract tests after #13724, #13749 and #13828 by @danielhanchen in #13846
Fix main's red CI after #13806, #13899 and recent formatter drift by @danielhanchen in #13909

New Contributors

@YashwanthNomula made their first contribution in #12647
@sahacha made their first contribution in #13606
@YogeshSeeni made their first contribution in #13783
@heychaeyin made their first contribution in #13813

Full Changelog: v0.1.907-beta...v0.1.908-beta

unslothai/unsloth v0.1.908-beta NVIDIA Nemotron 3.5 + Meta Muse Spark Open on GitHub