unslothai/unsloth v0.1.909-beta on GitHub

DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein are here! Run both locally in Unsloth. Also new: images and videos in chat, multi-GPU training with FSDP2 (alpha) and one-command MCP setup for coding agents.

Highlights

DeepSeek-V4.1-Flash-1017: 552B (16B active) - runs on ~125 GB at 1-bit with its Engram table on SSD Guide
FLUX.3 Image Klein: Runs on 24 GB and edits with up to 10 reference images GGUF
One-command MCP setup: unsloth mcp add connects Claude Code, Codex or Cursor to Unsloth.
Images and videos in chat: Local chat models use the image or video model loaded beside them.
Qwen-Image-Edit-2509 and 2511: Edit with up to 3 images at once.
Token preview: See what your model will learn before training starts.

DeepSeek-V4.1

Run DeepSeek-V4.1-Flash-1017, an update to DeepSeek-V4.1-Flash with better reasoning, coding and agentic results, and shorter outputs. It has 552B parameters plus 196B in its Engram table, about 16B active, and 1M tokens of context. With the Engram table left on SSD, it runs in about 125 GB of memory at 1-bit, or 300 GB at 4-bit. Guide · GGUF
DeepSeek-V4.1-Flash-1017 scores 42.3 on Terminal-Bench 4.0, up from 31.2 for the earlier DeepSeek-V4.1-Flash.
Thinking is on by default at high effort. Choose any effort from low to max, or turn thinking off.

FLUX.3 Image Klein

Generate and edit images with Black Forest Labs' FLUX.3 Image Klein, a 5.6B model that renders text and edits with up to 10 reference images. It runs on 24 GB of GPU memory, its GGUFs also run on computers without a supported GPU, and the undistilled Klein Base is available too. GGUF · Base GGUF
Enhance prompt, on by default, expands short prompts into detailed captions, which fixes most distorted bodies in hard poses. Preview shows the caption so you can copy or edit it.

Images + video

Qwen-Image-Edit-2509 and Qwen-Image-Edit-2511 now accept up to 3 images per edit.
Set custom video sizes and aspect ratios in Create videos, or pick one of the model's trained sizes.
Start an edit straight from an image's menu in the Images gallery or the Library, or drag a generated image onto an image picker on the Images page.
Image and video models with a large text encoder now load a quantized encoder by default where the GPU supports it, cutting the encoder's memory by up to half and loading faster. LTX-2.5 no longer needs a Hugging Face login.

Chat + Desktop

Image, video and Read aloud speech models can now stay loaded beside your chat model on computers with a dedicated GPU. Pick When they fit or Always in Settings > Resources > Model memory.
With the Images pill on, a local chat model can use the image or video model loaded beside it to make images and playable videos in the chat.
After an image or video model replaces your GGUF chat model, the chat continues without re-reading the conversation, as long as Keep chat context across unloads is on.
In Unsloth Desktop, choose an existing working folder when creating a project.
Unsloth Desktop no longer crashes at startup on older macOS versions, and its browser panel works on macOS 27 and follows system proxies such as Clash.

Training

In Python, spread LoRA, QLoRA and full fine-tuning across GPUs with FSDP2, and split long sequences, including packed ones, with context_parallel_size. Check the machine first with unsloth doctor --gpus.
On the Train page, Preview tokens shows which parts of your dataset the model trains on and flags samples cut off by the context limit. Data Recipe chats now train as multi-turn conversations, and Qwen fine-tunes keep a dataset's thinking text.
Cut full fine-tuning optimizer memory with optim = "adamw_flash", or set UNSLOTH_OFFLOAD_OPTIMIZER_STATE=1 to keep optimizer state in system memory between steps at some cost in speed.

Performance

Local GGUF chats with tools now show replies as fast as the model writes them, so the next tool call no longer waits, and long replies no longer slow down as they grow.
LoRA training with lora_dropout or lora_bias now runs on Unsloth's fused kernels, about 28% faster per step.
Gemma 4's global attention layers train on a new kernel on NVIDIA Ampere or newer GPUs, with 1.4x to 2.1x faster attention and up to 4.9x less attention memory.
Qwen3.8-Flash-Next on MLX is up to 1.4x faster on M5 Macs and up to 1.2x faster on M1 to M4 Macs.
On M1 to M4 Macs, 4-bit and 8-bit MLX models generate several chats at once up to 1.8x faster.

Coding agents

Connect Claude Code, Codex or Cursor to Unsloth's MCP server with unsloth mcp add claude-code (or codex, cursor).
Agents can upload datasets, start training and load the resulting checkpoint.
Agents can also download models ahead of time, preview what a model load will download or unload, and queue video generation jobs.

Models + settings

Filter downloaded models by capability or format, including Exports & Fine-tunes, and search for a maker such as DeepSeek or Meta to list Unsloth's uploads of its models.
QwenCloud and Z.ai are now available as API providers.
Export token usage from Profile stats as CSV.
unsloth run passes llama-server short flags such as -ctv, -dev and -ctxcp through unchanged.

Hardware compatibility

PCs without an NVIDIA GPU or ROCm, such as Snapdragon laptops and Linux ARM boards, now run GGUFs on their GPU through Vulkan instead of the CPU.
NVIDIA MIG slices are sized and selected correctly, and machines that mix GPU generations get settings and llama.cpp builds every GPU can run.
On Linux machines with an AMD Strix APU and a discrete AMD GPU, PyTorch is installed for the discrete GPU. On Windows AMD PCs with the HIP SDK, Settings > System no longer asks for administrator access.
Windows on ARM PCs can now turn on Sandbox High in Settings > Sandbox.

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

Bump install.sh / install.ps1 pins to unsloth>=2026.10.6, unsloth-zoo>=2026.10.6 by @danielhanchen in #13918
Studio: preview tokens and the completion mask before training by @danielhanchen in #13850
Locate the MTP drafter-drop fallback by its shape, not one spelling by @danielhanchen in #13922
Fix main's red Backend CI after #13850 and #13806 by @danielhanchen in #13930
Studio: do not re-send the image generate POST after a dropped connection by @drakeo338 in #13790
NVFP4: decode weight tiles inside the GEMM for small-batch forward and dX by @danielhanchen in #13840
Fix main's red CI: UI font tokens in the token preview, real-CUDA skip guards by @danielhanchen in #13936
NVFP4 tests: real-CUDA skip guards, and format the files #13840 left off the hook by @danielhanchen in #13937
Studio: prefer NVIDIA Wayland dmabuf over static compositing-off by @indrajeetapache in #13892
Baseline the unsloth-zoo 2026.10.6 and transformers 5.20.0 scan findings after review by @danielhanchen in #13949
Add context parallelism support (SDPA only) by @danielhanchen in #4257
Prebuilt CUDA wheels: pin NVIDIA's re-published CUDA 13.4.1 manifest (win_arm64) by @danielhanchen in #13954
GRPO: keep TRL's old logprob gate when using vLLM by @danielhanchen in #13856
Refuse an explicit offload_layers swap under FSDP / DeepSpeed by @danielhanchen in #13857
RMSNorm: opt-in saving of the norm output instead of its input (UNSLOTH_RMSNORM_SAVE_OUTPUT) by @danielhanchen in #13868
Add a Triton head_dim 512 attention kernel for Gemma-4 global layers by @danielhanchen in #13869
Use side stream offload hooks for non-reentrant gradient checkpointing by @danielhanchen in #13872
GRPO: opt-in parking of optimizer state and gradients on host during rollout by @danielhanchen in #13866
Studio: automatic image/video loads pick the freest GPU, and GPU picks respect what each card can run by @danielhanchen in #13928
Studio installer (Windows AMD): treat hipinfo output as already mask-filtered when picking the gfx target by @danielhanchen in #13925
Context parallel: Ulysses backend with packing and varlen attention by @danielhanchen in #13848
GRPO packed logprobs: an OOM caps the packed length instead of disabling packing by @danielhanchen in #13877
Build RoPE cos/sin caches only on the device that uses them by @danielhanchen in #13855
Opt-in token-balanced and token-budget SFT micro-batches by @danielhanchen in #13878
GRPO: exclude masked rows from clip_ratio metrics for one-token completions by @danielhanchen in #13883
Studio installer: cover the GPUs CUDA_VISIBLE_DEVICES hides when picking the llama.cpp CUDA bundle by @danielhanchen in #13924
Context parallel: Ulysses for Gated DeltaNet hybrids (Qwen3.5, Qwen3-Next) by @danielhanchen in #13920
Core: decide bfloat16, compile and TMA options from every visible NVIDIA GPU on mixed hosts by @danielhanchen in #13923
Studio: the bot-check page's Unsloth Desktop link opens Get started by @shimmyshimmer in #13958
Fix RMSNorm and LayerNorm backward overwriting the incoming gradient by @danielhanchen in #13888
DPO, ORPO, CPO: opt-in fused preference loss (UNSLOTH_FUSED_PREFERENCE_LOSS=1) by @danielhanchen in #13870
Add opt-in optimizer state CPU offload (UNSLOTH_OFFLOAD_OPTIMIZER_STATE=1) by @danielhanchen in #13882
Studio: Deep research sits under Tools in the + menu by @shimmyshimmer in #13959
fix(studio): load audio GGUFs selected from custom folders by @Imagineer99 in #13897
Studio: name the open connection, lead the provider menu with Custom, new OpenRouter logo by @shimmyshimmer in #13964
FSDP2: first-class LoRA and full finetuning, keep fused LoRA kernels per block by @danielhanchen in #13865
Studio: proportional image attachments, attachment card and Chat with files polish by @shimmyshimmer in #13970
Installers (AMD, Linux): a Strix APU defers torch to a discrete card, and a lone gfx1153 warns instead of getting kernel-less wheels by @danielhanchen in #13931
Keep the plain capability cache on single-GPU hosts by @danielhanchen in #13962
Add optim = "adamw_flash" (FlashAdamW from flashoptim) for lower optimizer memory by @danielhanchen in #13881
Fused Qwen3 q_norm / k_norm RMSNorm + RoPE Triton kernel (opt-in) by @danielhanchen in #13890
GRPO: opt-in zero-variance group exclusion and prompt-level loss aggregation by @danielhanchen in #13891
Tests: gate the optimizer state offload CUDA cases on a real CUDA device by @danielhanchen in #13982
Studio: raise the startup bundle budget by the growth since 5870795 by @danielhanchen in #13983
Studio: let safetensors and MLX vision models read transparent images by @NilayYadav in #13939
Studio: honour MIG masks and size MIG cards by their slice by @danielhanchen in #13927
Fix optimizer state offload crash with Q-GaLore weight quantization by @danielhanchen in #13979
Context parallel: compose with FSDP2 on a (dp_shard, cp) mesh by @danielhanchen in #13910
Studio: train a structured column's conversation as a multi-turn chat by @NilayYadav in #13940
Studio: line the composer up with the chat column in the macOS Desktop app by @shimmyshimmer in #13965
Studio: fix the context meter after tool calls on MLX and transformers by @NilayYadav in #13913
Studio: resizable Model hub split view, smaller row kebab, centred token icon by @shimmyshimmer in #13966
Studio: rank Latest Unsloth Models by trending weighed against release age by @shimmyshimmer in #13969
Studio: draw chat markdown tables inline with hover controls and a centered expand dialog by @shimmyshimmer in #13973
Studio: Models panel part of its card, even form footer, OpenRouter mark under a new file name by @shimmyshimmer in #13974
Studio: even 17px spacing in the Hub filter popover by @shimmyshimmer in #13972
Sync full finetune weights into colocated vLLM in a few load_weights calls by @danielhanchen in #13901
Studio: keep the tool results of the turn in progress when compaction resets mid-turn by @oobabooga in #13951
GRPO: keep TRL vLLM weight sync on TRL < 0.28 unless fast_inference shares weights by @danielhanchen in #13911
RMSNorm save output: skip inference tensor weights, keep the default path free of opt-in checks by @danielhanchen in #13981
Studio: hide safetensors models that exceed VRAM when the fit filter is on by @oobabooga in #13945
GRPO no-grad packing: check the OOM length cap before flattening ids by @danielhanchen in #13980
FSDP2: shard QLoRA 4-bit weights with fused LoRA kernels kept on by @danielhanchen in #13880
FSDP2: save_pretrained, merged saves and 4-bit checkpoints for sharded models by @danielhanchen in #13895
Add unsloth doctor --gpus multi-GPU preflight and UNSLOTH_DEBUG_DISTRIBUTED by @danielhanchen in #13884
GRPO: latch the trainer's autocast precision on the efficient logprob path by @danielhanchen in #13968
Studio: raise the startup bundle budget by the growth since 1790106 by @danielhanchen in #13991
Studio: Chat with files X hides the panel, app tooltips, and higher-contrast traces by @shimmyshimmer in #13989
head_dim 512 attention: sm80+ only, no 256 MB autotune buffer, OOM falls back to SDPA, no per-call mask sync by @danielhanchen in #13977
Tests: gate the new Q-GaLore offload CUDA case on a real CUDA device by @danielhanchen in #13994
Studio: raise the startup bundle budget by #13989's growth by @danielhanchen in #13995
Studio: scale the expanded markdown table's text with the UI font size by @danielhanchen in #13996
Cover explicit RoPE position ids and compute the starting cache on CPU by @danielhanchen in #13976
Studio: a hidden Chat with files panel stays hidden after the first message by @shimmyshimmer in #13999
Studio: model picker ranks Recommended by the Model hub's Latest ranking by @shimmyshimmer in #13993
Studio: searching a model maker in the Model hub lists Unsloth's uploads of its models by @shimmyshimmer in #13998
Token budget: weight split-chunk TRL metrics by token count by @danielhanchen in #13978
Add opt-in TFLOPs and MFU logging with LoRA aware FLOP counts by @danielhanchen in #13885
Studio: raise the startup bundle budget by #13998's growth by @danielhanchen in #14004
PEFT MoE conversion snapshot: add beam, which transformers main now fuses as qwen2_moe by @danielhanchen in #14005
studiobench CI: boot Studio with update checks off so a mid-run llama.cpp release cannot cover the scene by @danielhanchen in #14007
Fused LoRA kernels: support lora_dropout and lora_bias by @danielhanchen in #13898
RMSNorm: restore in-place dX for pre-norm decoder norms by @danielhanchen in #13984
Keep opted-in bnb 4-bit MoE experts outside the compiled block by @danielhanchen in #13946
Studio: improve AMD NPU discovery, CLI support, and chat feedback by @oobabooga in #13797
Studio: faster frontend type check with TypeScript 7 and bun check by @Etherll in #13986
GRPO: score old and reference logprobs over the truncated sampling support when opted in by @danielhanchen in #13900
GRPO: install MoE routing replay when UNSLOTH_GRPO_ROUTING_REPLAY=1 by @danielhanchen in #13916
Kernel follow-ups: norm launch on HIP and under compile, grouped GEMM frozen operands, no CE eviction hint on H100 by @danielhanchen in #13967
Windows hipinfo: an all-hiding ROCR mask still means no ROCm target by @danielhanchen in #14019
Optimizer state offload: keep inactive state on the host for unchunked Q-GaLore weight_quant steps by @danielhanchen in #14025
perf(studio): enter Zoo's inference scope registry on Unsloth Studio's MLX paths by @Lyxot in #13975
Move the Unsloth Studio MLX stack to mlx 0.32.4 and mlx-lm 0.32.2 by @Lyxot in #13992
FP8 GEMV on Blackwell + Triton 3.3: no graph break under torch 2.7, keep the exact GEMV shapes that are faster by @danielhanchen in #14026
NF4 GEMV: keep the table load on 256 wide tiles by @danielhanchen in #14027
Context parallel: gather the FSDP2 state dict in save_model on transformers 4.x by @danielhanchen in #14028
Follow-ups to the GRPO reward locals and FSDP2 mixed precision merges by @danielhanchen in #14023
Token budget: weight split entropy by labels on TRL 1.3 to 1.14 by @danielhanchen in #14024
AMD: Strix defers only to an installable generic tag; gfx1153 health gate classified at inventory refresh by @danielhanchen in #14022
Studio: list every GGUF build that shares a quant in a custom folder by @oobabooga in #13957
Studio: custom video sizes and Images-style prompt fields on Create videos by @shimmyshimmer in #14050
Studio: show a finished chat's sidebar dot in the accent colour by @shimmyshimmer in #14042
Studio Hub: sidebar-matched pill icons, drop the Memory stat by @shimmyshimmer in #14049
Studio: API page toolbar gets clearer icons and hover tooltips by @shimmyshimmer in #14048
Studio: read whole-GPU uuid masks from nvidia-smi rows and cache the MIG set by @danielhanchen in #14017
Studio: keep automatic GPU placement inert on ordinary hosts by @danielhanchen in #14020
Chat UI smoke: boot Studio with update checks off by @danielhanchen in #14053
CUDA spoof: cover CUDA wheels on GPU-less runners, not only CPU wheels by @danielhanchen in #14056
Notebooks smoke: run the install cell the way Colab does by @danielhanchen in #14058
Studio: align toast icons with the first line on multi-line toasts by @shimmyshimmer in #14082
Studio Hub: On Device filters and sort, Exports & Fine-tunes, unsloth.ai links by @shimmyshimmer in #14083
Studio: slide the selected pill on every segmented toggle by @Etherll in #14030
Studio: use the inline markdown table style in the Hub, file previews and the composer draft by @shimmyshimmer in #14087
Studio: browser panel follows the system proxy on macOS, Windows and Linux by @Lyxot in #14029
Studio: a managed account's Hub dataset card may not make the server dial an http(s) URL by @danielhanchen in #14091
CI: pin the Arrow and sqlite-vec tags the Windows ARM wheelhouse signs to commits by @danielhanchen in #14092
Tests: make the --help, media scan, installer comparer and harness resolver checks hold on every host by @danielhanchen in #14116
Studio: automatic GPU placement caches failed probes, skips MIG winners, never blocks the event loop by @danielhanchen in #14118
Studio: tell Gemma 4 with Think off and GLM today's date by @NilayYadav in #14013
Let the Gemma-4 chat template format tool-calling conversations by @NilayYadav in #14015
Create the branch before pushing a LoRA GGUF to a revision by @NilayYadav in #13915
Studio: keep skills created in Docker on the unsloth-studio volume by @NilayYadav in #13912
Studio: fit knowledge base chunks to the embedding model's input limit by @NilayYadav in #14045
Studio: keep a chat's files in the Library after it moves into a project by @NilayYadav in #14044
fix(studio): guard missing scrollbar-gutter on Mac desktop startup by @Souravrajvi0 in #14111
Studio: read uploaded CSV cells as written for embedding training by @NilayYadav in #14014
Studio: push an export to the Hub with the Hugging Face CLI login by @NilayYadav in #14047
Studio: keep a dataset's thinking text when training Qwen models by @NilayYadav in #14016
Studio: name unknown GGUF quant types below the bundled type count by @danielhanchen in #14124
Studio: let the browser panel and web fetch work behind fake-IP DNS proxies by @danielhanchen in #14127
Studio: let a finished reply be edited when its settle write was refused by @danielhanchen in #14131
Unsloth Studio / Desktop (AMD, Windows): stop amd-smi from popping UAC/DiskPart prompts on the System page by @LeoBorcherding in #14041
Load six older models in 4bit instead of a missing repo by @NilayYadav in #13941
Studio: skip a Terminal PATH entry MXC cannot grant instead of failing every launch by @danielhanchen in #14129
Studio: let the Project sources panel be hidden in a project chat by @danielhanchen in #14125
Model picker contract: follow the On Device filter onto Discover's capability picks by @danielhanchen in #14138
Studio: keep GGUF variant listings off the default thread pool so a stalled Hub cannot block downloaded models by @danielhanchen in #14134
Studio: unload a local model without stopping external provider chats by @NilayYadav in #13917
Studio: keep llama.cpp and vLLM connections' default system prompt by @NilayYadav in #14043
Studio: default to the Vulkan llama.cpp bundle when a CPU-bound host has a Vulkan GPU by @danielhanchen in #13599
Studio: name a closed AMD device node in the System tab GPU notice by @danielhanchen in #14126
Studio: ask before risky calls when a chat folder is too large to check by @NilayYadav in #13914
Fix Nemotron loss normalization across gradient accumulation by @taking-lying-flat in #14002
Kernel follow-ups: NVFP4 shape-aware fused gate, MoE fallback tiles, test fixes across torch 2.10 - 2.14 by @danielhanchen in #14122
Studio: wrap long commands and cap tall API examples in Settings by @Etherll in #14071
Studio: ask before ffmpeg or pandoc write outside the sandbox by @NilayYadav in #13942
Read the current FSDP2 output head when scoring GRPO log-probs by @taking-lying-flat in #14060
Studio: add QwenCloud as an API provider by @NilayYadav in #14033
Fix LM-head weight fake quantization in fused CE by @taking-lying-flat in #14006
Load Ling 3.x low-rank queries on native bailing_hybrid and fix its exports by @danielhanchen in #14136
Fix packed document boundaries in Qwen PLE by @taking-lying-flat in #14011
Studio: FLUX.3 Image Klein (distilled and base), with INT8 / FP8 / GGUF support by @danielhanchen in #13933
Studio: install from a package index that lags PyPI by @danielhanchen in #14133
Studio: add Z.ai as an API provider by @NilayYadav in #14034
Studio: keep Claude's web search to the sites it asks for by @NilayYadav in #14046
Studio: audio.cpp v0.9.2 support by @Etherll in #14073
Studio: FLUX.3 Image Klein on the native sd.cpp engine by @danielhanchen in #14144
Studio (macOS): size image loads against the Metal working set, not RAM minus 20% by @oobabooga in #13776
Studio: reserve MiniMax-H3's pinned weights from sd.cpp's budget on integrated AMD GPUs by @oobabooga in #13786
Studio: video families default to a hosted quantized text encoder by @danielhanchen in #14000
Studio: FLUX.3 prompt upsampling (Enhance prompt) through a llama.cpp sidecar by @danielhanchen in #14137
Studio: fetch every image component from unsloth repos by @danielhanchen in #14008
Studio: image families default to a hosted quantized text encoder by @danielhanchen in #14009
Studio: fetch every video component from unsloth repos by @danielhanchen in #14003
RMSNorm: keep the final norm's own dX buffer when hidden states are returned to the caller by @danielhanchen in #14139
AMD install follow-ups: installable inferred target, gfx1153 lspci peer, override health hint by @danielhanchen in #14117
Log budget: measure the FLUX.3 prompt upsampler status poll by @danielhanchen in #14167
Tests: stop a GPU-selection route test leaking a CHAT owner into the upsampler tests by @danielhanchen in #14169
Studio: keep the download button off the chat composer on phones by @shimmyshimmer in #14163
Studio: audio.cpp models land on the right pages by @Etherll in #14072
Studio: one-command MCP setup, recipe prompts, a model catalog and tool annotations by @Etherll in #14077
Studio: clearer audio.cpp errors, cleaner Speak/Clone, reproducible takes by @Etherll in #14075
Studio: image and video generation follow each model's defaults and prompting tips, and /v1/videos takes Unsloth Studio controls by @Etherll in #14078
Studio: drop an Images strip tile onto an image picker to use it by @shimmyshimmer in #14166
Studio: add Edit image to the Images gallery item menu by @shimmyshimmer in #14162
GRPO: latch autocast under vLLM only, zero-variance exclusion on TRL before 1.6 by @danielhanchen in #14123
Studio: add a setting to stop chats and Deep Research when the tab closes by @alkinun in #14062
Studio: MCP audio reports what it applied, explains audio errors, checks takes and cancels by id by @Etherll in #14080
Studio: the Images pill makes images and videos in chat with the loaded local models by @Etherll in #14110
Studio: MCP returns key-free download links and finds media, voices and video frames again by @Etherll in #14079
Studio: add an Always choice to Keep multiple models loaded, track every model on the GPU, and report every model a load unloads by @Etherll in #14088
Studio: image and video models stay loaded beside chat when they fit by @Etherll in #14106
Studio: MCP errors, cancels and chat controls agents can act on by @Etherll in #14081
Studio: MCP agents follow long work in one call and see every model a load unloads by @Etherll in #14089
Studio: choose Off, When they fit or Always in Model memory, and keep the chat context when an image or video model takes the GPU by @Etherll in #14108
Studio: show the running step on the Hub card's download bar by @shimmyshimmer in #14171
Studio: MCP agents choose how and where a model loads, preview the load, and queue videos by @Etherll in #14090
Studio: MCP agents upload datasets, load what they trained, and download models ahead of time by @Etherll in #14107
Studio MCP: keep one [code] tag on audio errors by @danielhanchen in #14175
Studio: add Edit image to Library image menus by @shimmyshimmer in #14170
Studio: load the retired FLUX.3 full copies through their base by @danielhanchen in #14173
Studio: download the UD-Q4_K_XL FLUX.3 prompt upsampler by default by @danielhanchen in #14177
Studio: label the Chat with files composer pill Files instead of RAG by @shimmyshimmer in #14179
Studio: match letter shortcuts by the letter the layout types by @Souravrajvi0 in #14113
Desktop: keep DMA-BUF off on NVIDIA Wayland when compositing is turned off by the environment by @danielhanchen in #14141
Studio: darker unselected model picker toggle, collapse icon for Compact conversation by @shimmyshimmer in #14183
CLI: doctor --mcp sends the API key only to a Studio that proves its identity by @danielhanchen in #14178
Studio: run the Windows sandbox (MXC) on ARM64 PCs with Microsoft's ARM64 build by @danielhanchen in #14176
Studio: one tooltip design everywhere, matching chat's text colours in light and dark by @shimmyshimmer in #14172
Studio: pin FLUX.3 Image Klein (Fast FP8) to the top of the Images picker by @danielhanchen in #14184
Desktop: bump rustls to 0.23.46 and quinn-proto to 0.11.19 for RustSec advisories by @danielhanchen in #14186
Studio: keep the prompt upsampler client out of the startup chunks, reset the startup budget to the build by @danielhanchen in #14168
Studio: update Lemonade to 2026.42.1 and FastFlowLM to v1.0.9 by @danielhanchen in #14189
Notebooks CI: refresh the Colab oracle snapshots to the current GPU runtime by @danielhanchen in #14182
Studio: managed accounts use the server's own HF endpoint for recipe seeds by @danielhanchen in #14187
Studio: open //host links as web links; run apt-get by absolute path; ignore the dev Studio home by @danielhanchen in #14190
Studio: raise the startup bundle budget for #14190 by @danielhanchen in #14198
Studio: keep the FLUX.3 Image Klein GGUF row under its Fast FP8 row at the top of the Images picker by @danielhanchen in #14197
Studio: strip tool markup from streamed text without rescanning the whole turn per token by @danielhanchen in #14143
Studio: raise the startup bundle budget for #14197 by @danielhanchen in #14203
Hybrid packing + block compile: fix the guard crash, the skipped varlen, and the cross-document leak by @Etherll in #13956
Studio: compile FLUX.3 Klein once for every prompt length by @danielhanchen in #14211
Copy the residual before the in-place norm in Granite fast decode by @vineethsaivs in #13454
Studio: follow a same-model MLX reload in the KV width and int8 prefill settings by @danielhanchen in #14210
Studio: install the S3 SDK on demand by @AlbertoLMa in #14157
Studio: start the FLUX.3 prompt upsampler while Enhance prompt is on by @danielhanchen in #14208
Fix padded prefill sliding window being one key too wide by @vineethsaivs in #13906
Studio: fix browser toolbar alignment and narrow layouts by @Imagineer99 in #14018
GRPO: file compute_loss metrics by model.training, not control.should_evaluate by @vineethsaivs in #14112
Studio: keep users signed in when auth.db is busy by @Etherll in #14151
Fused QKNorm + RoPE: keep type promotion for mixed dtype tables, reuse grad in the trainable norm fallback by @danielhanchen in #14200
Studio: export token usage from Profile stats as CSV by @danielhanchen in #14213
Studio: keep reading durable chat output while its events are being saved by @danielhanchen in #14135
unsloth run: pass llama-server short flags like -ctv, -dev, -ctxcp through whole by @danielhanchen in #14224
Context parallelism: drop the SDPA wrappers each ring step leaves in torch's registry by @danielhanchen in #14193
head_dim 512 attention: use the kernel for masks made under inference mode by @danielhanchen in #14202
Pass layers_to_transform, use_rslora and modules_to_save to LoraConfig on the SentenceTransformer fast encoder path by @vineethsaivs in #14150
fix(studio): reject oversized Metal alongside loads before startup by @Imagineer99 in #14209
Block compile: keep the GatedDeltaNet self-check out of a non-reentrant checkpoint by @oobabooga in #14195
Studio: multi-image editing for Qwen-Image-Edit by @danielhanchen in #14230
Studio: fix chat scrollbar freezes on long code fences by @Sneakr in #13919
Studio: restore Qwen causal-conv1d on demand, narrow fatal mamba-ssm, provision SSM kernels in export by @danielhanchen in #13971
Attention: GQA never forces SDPA math, cuDNN-over-varlen on sm100 only, native before xformers by @danielhanchen in #14158
Prebuilt wheels: drop -lineinfo, unreachable SASS and flash-attn PTX; add extension-free mamba-ssm by @danielhanchen in #14142
Studio: stage prebuilt installs in dirs that inherit the parent ACL on Windows by @danielhanchen in #14212
Optimizer state offload: track DTensor state loaded on the host when resuming by @danielhanchen in #14155
doctor --gpus: read NVML P2P enum constants as ints by @danielhanchen in #14164
RoPE: build grown caches on the GPU again, 32768 rows at a time by @danielhanchen in #14206
FP8 GEMV long-indexing test: seeded inputs and a last-bit tolerance by @danielhanchen in #14220
Studio: choose working folders when creating projects by @Imagineer99 in #13873
Qwen3 decode: size the RoPE read by the explicit position, not only the KV length by @danielhanchen in #14156
Studio: keep the Read aloud speech model loaded beside the chat model by @Etherll in #14191
Studio: keep code colour while scrolling long chats by @danielhanchen in #14242
Studio: route clicks and scrolls to the browser page on macOS 27 by @danielhanchen in #14216
pre-commit: cut the kwarg formatter's CPU 6.5x and stop the ruff fork bomb (#14174) by @danielhanchen in #14249
Studio: build the OpenAPI schema off the event loop by @danielhanchen in #14239
FSDP2 QLoRA: give the all-gathered 4-bit weight back its Params4bit class and 3D shape by @danielhanchen in #14204
Studio: Developer mode opt-in for experimental CPU video generation by @danielhanchen in #14247
Studio: raise the startup bundle budget for #14247 by @danielhanchen in #14258
Honor layers_to_transform and finetune_last_n_layers with regex target_modules by @vineethsaivs in #14021
feat(studio): add a setting to hide the macOS menu bar icon by @Mathews-Tom in #11937
CI: keep one generation of prefix-restored caches, share Core's pip cache by @danielhanchen in #14263
Studio: delete selected recipe blocks with the Delete key by @danielhanchen in #14223
Studio Images: per-load engine choice for GGUF picks (native sd.cpp for LoRA) by @danielhanchen in #14232
Studio: wire in DeepSeek-V4.1-Flash-1017 (deepseek41) by @danielhanchen in #14260
Studio: raise the startup bundle budget for #11937 by @danielhanchen in #14264
Studio: make inpaint and localized-edit masks from text with SAM 3 by @danielhanchen in #14235
Studio installer: show the previous unsloth version next to the new one by @danielhanchen in #14248
Studio: show the end of long prompts in the API monitor by @danielhanchen in #14241
Context parallel: wrap trainer methods only once a trainer enables CP (torch 2.10 step time) by @danielhanchen in #14152
Studio: a kept load short on its own GPU no longer unloads models on other GPUs or on the CPU by @danielhanchen in #14234
Decode: cover explicit positions past the lazy RoPE cache in Gemma2, Cohere, Granite and FalconH1 by @danielhanchen in #14199
MFU logging under fullgraph, NCCL logs under distributed debug, token-budget metrics across ranks by @danielhanchen in #14207
Studio: let an approved rm run, never a recursive one by @danielhanchen in #14238
Studio: eject a model with a running chat from the Loaded models card by @NilayYadav in #14161
Unsloth Studio / Desktop: name chat exports after model, title and day, and reopen the last save folder by @LeoBorcherding in #12586
Tests: a grandchild reaped while its /proc stat is read counts as gone by @danielhanchen in #14268
Tests: echo gguf_filename in the image text encoder Playwright status stub by @danielhanchen in #14269
Tests: make image-engine-choice and image-mask-from-text typecheck by @danielhanchen in #14272
Studio: timestamp every other sampled video frame for GGUF chat by @danielhanchen in #14254
Format the two files #14247 left off the formatter fixed point by @danielhanchen in #14270
unsloth start vscode: print the group instead of writing a dead file inside a plain container by @danielhanchen in #14237
Studio: a chat remembers the model it last used, not only the one it started on by @danielhanchen in #14259
Tests: start every diffusion_segment test with no GPU arbiter resident by @danielhanchen in #14271

New Contributors

@AlbertoLMa made their first contribution in #14157

Full Changelog: v0.1.908-beta...v0.1.909-beta

unslothai/unsloth v0.1.909-beta DeepSeek-V4.1 & FLUX-3 on GitHub