Unsloth Release Notes

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.

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.

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

Turn any text or vision LLM into a Jev-style decision model in Unsloth, with decision accuracy going from 30% to 80%. Train, test, export and serve decision models directly from Unsloth. Also included: native ComfyUI models, diffusion improvements and a better Browser in Desktop.

Highlights

Turn any model into a Jev-style decision model. Accuracy 30% to 80%Load ComfyUI diffusion models natively in UnslothFaster + more accurate diffusion with INT8 ConvRot and moreBetter Browser in Desktop with many bug fixesSandboxing with Bwrap for Linux, Seatbelt for Mac and MXC for Windows

This release adds a browser inside Unsloth (browser use coming very soon), so files, web pages and pages the model writes open right beside your chat. It also brings EmbeddingGemma 2, Google's new multimodal embedding model, plus new Audio pages and lower-memory training.

Highlights

Browser beside chat for files, web pages and model-written HTML, as tabsEmbeddingGemma 2: run and fine-tune Google's new model for text, code, images, video and audio GuideAudio pages: Speak, Clone, Music, Transcribe, Edit, Separate, Convert, with 70+ GGUF modelsKeep multiple models loaded, plus vLLM and SGLang engines (opt-in)Sandbox for code the model runs, now with a Settings tab and Colab and Docker supportLive preview during image and video generation

This release makes Unsloth Desktop faster and easier to use with improved navigation, shareable run settings, clearer errors, faster Laya decisions, expanded Decision API support, and better 4-bit LoRA training.

Desktop, chat, and models

Cmd/Ctrl+P command palette and shareable GGUF run settings. Laya decisions up to 4.1x faster with hosted provider support. Clearer errors with improved load/generation messages and logs. Lower-memory 4-bit LoRA training for NVFP4, INT4, and MXFP4 checkpoints.

Share GGUF run settings without loading models. Continue responses and fix HTML canvas errors directly in chat. Improved model fit warnings, image loading, and desktop controls. Added FastModel fine-tuning support for T5, T5Gemma, BART, Marian, and more. Added unsloth eval for evaluating checkpoints and LoRA adapters. Improved multi-GPU training and generation support.

4-bit checkpoint training

Pre-quantized checkpoints now train using their original packed weights, reducing memory use while preserving accuracy.

Qwen3.8-27B-NVFP4 uses 40.2 GB peak memory instead of 72.9 GB. INT4 checkpoints stay packed instead of being re-quantized. Kimi-K2.7-Code and gpt-oss retain packed experts during LoRA training.

Decision API, image, and platform updates

Laya decision models respond up to 4.1x faster, with hosted providers available through Connections. Image models preserve INT8/FP8 during offloading for faster generation. Qwen-Image-2.1 renders 1536×1536 images up to 3.6x faster on Radeon 8060S. Added AMD Windows support for RX 5000–9000 GPUs. Updated MLX support and Mac LoRA exports to PEFT or GGUF.

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

Run and serve Decision Models like Laya (open-source Jev) locally through a TypeSafe-compatible API.

Skills Editor to create, edit, and delete Skills directly in Desktop.

ModelScope model downloading is now available for users who can't use Hugging Face

Document Viewer for PDF, Word, Excel, PowerPoint, and chats.

New Library: Manage chats, images, videos, and files from the new Library tab.

Apple Silicon Improvements including batched serving, structured outputs, and TurboQuant KV cache.

Configure how much of a model each GPU receives.

Faster Image + Video Generation with ~4.5x faster LTX-2.3 clips and 1.7–6.3x faster VAE decoding.

You can now run Qwen-Image-2.1 locally with Unsloth! This release also includes custom Agent Skills, and easier chat/project management. It also brings 2x faster reasoning blocks (60 FPS vs 30 FPS), more reliable training, and improved Linux installs and updates.

Qwen-Image-2.1 support

Run Qwen-Image-2.1 locally with Unsloth for image generation.

Custom Agent Skills

Add task-specific skills, reuse skills from Claude Code and .agents folders, and select them in chat with @.

Smoother chat experience

Long reasoning blocks now render at 60 FPS, up from 30 FPS. Includes a redesigned thinking UI, smoother code streaming, and more reliable model loading and chat settings.

Easier chat and project management

Drag chats to reorder, pin, or move them into projects across all platforms. Edit project names, instructions, and folders directly from the Projects page.

Better model and storage handling

Vision models can see images returned by MCP tools. Reuse downloaded models without fetching duplicate copies, and inspect or clear caches from Settings.

More reliable training and exports

Resume past training runs with accurate remaining-time estimates. Fixes improve Hugging Face dataset recipes, column mapping, and split/subset selection. Hub uploads exclude files left over from earlier exports.

Improved Linux and AMD support

In-app updates for Debian installs, a native ARM64 installer for Ubuntu 24.04+, and a new AMD ROCm Docker image with Unsloth Studio and JupyterLab.

We're releasing our new updated Docker image along with multi-user accounts, RDNA1+2, FP8/INT8 diffusion support, ARM64 CUDA Windows support and many training, GRPO and inference improvements

Highlights:

• Updated Docker with NVIDIA and AMD support: Guide
• Multi-user accounts with isolation: Settings > Accounts
• INT8/FP8 image diffusion inference: 2x faster
• ARM64 Windows CUDA support for training and inference
• GRPO: Qwen3.5 and latest TRL/vLLM support
• AMD RDNA1 and RDNA2 support
• Better Windows NVIDIA GPU detection and recovery

Updated Docker image

Updated CUDA image and Studio setup, removed bundled caches, and restored Unsloth’s training patches on GPU hosts. Studio data persists on a volume without pinning app code. Improved updates, generated passwords, configurable ports, and package preservation. Added an AMD ROCm image for supported Linux hosts.

Multi-user accounts

Create accounts in Settings > Accounts with one-time setup codes and individual passwords. Accounts keep work separate while sharing a loaded model when settings match. Single-account behavior is unchanged.

Chat + reasoning

Edit, reorder, and steer queued prompts, even while a local model loads. Added GGUF reasoning budgets, adjustable chat width, desktop scaling, and a compact composer. Improved long-reasoning responsiveness, chat-history preservation, and tool-generated file and image handling.

Hardware + inference

MLX gains video input, optional MoE and decode optimizations, and more reliable multimodal chats. Improved GGUF memory estimates, DGX Spark handling, NVIDIA detection, NVLink checks, and user-selected GPU ordering. Added native Windows ARM64 desktop packaging, Qwen3.5 fast_inference, and Ascend NPU detection.

Research + API

Deep Research better preserves unfinished reports, handles JSON output, and flags runs without evidence. Scanned PDF uploads gain a local OCR fallback. Improved provider reasoning controls, MCP compatibility, tool calling, and streamed error reporting. Added /v1/chat/completions video input and API access to installed Ollama models.

Training + installation

Added Data Recipes dataset downloads and fixed 16-bit fine-tune exports, LoRA saves, and Push to Hub. Improved dataset handling, training memory estimates, and checkpoint saves on container shutdown. Core fixes cover normalization, RoPE, Q-GaLore, and distributed device selection. Improved Windows installs, GPU package repair, dependency caching, and verified offline installs.

Security + tool execution

Strengthened credential protection and tool-call validation, with approval requests for file access outside the sandbox. Added preview OS-level tool sandboxing on supported Linux and macOS hosts. Network access remains unrestricted; automatic mode keeps existing safeguards when OS isolation is unavailable.

Run Qwen3.8-Flash-Next and GLM-5.3-Flash up to 2× faster with MTP, now enabled by default. This release also includes 170+ improvements across training, chat, hardware support, APIs, and performance.

Faster Qwen and GLM generation with MTP, plus improved tool use and recommended Qwen settings

Much less laggy UI/UX for all chats especially on long conversations

New support for MiniMax-Music3, Higgs, and MOSS audio models, including progress tracking and clip management.

More reliable model loading across local servers and connected providers.

Safer chat editing that preserves tool calls, replies, media, and conversation branches.

Improved multi-GPU training, memory fitting, model placement, and GGUF exports.

Stronger AMD/ROCm installation, detection, and GPU compatibility.

More reliable MCP, Deep Research, OAuth, parallel tool calls, and agent workflows.

New OpenAI-compatible APIs for video, audio, and MLX-served models.

Improved desktop updates, LAN port controls, downloads, and GPU memory cleanup.

Qwen3.8-Flash-Next and GLM-5.3-Flash can now run locally in Unsloth!

Qwen3.8-Flash runs on 75GB RAM
GLM-5.3-Flash runs on 102GB RAM + VRAM
Up to 5x faster RAM offloading
Repeated Auto Compaction now works reliably
100+ chat, reliability and performance improvements

Qwen3.8-Flash-Next

125B multimodal reasoning model and early preview of Qwen4's architecture.

1-bit Dynamic GGUF runs on 75GB RAM or unified memory
Up to 262K context with text and images
None, Low, Medium and Extra High reasoning modes
Preserved Thinking improves consistency in long chats

GLM-5.3-Flash

320B multimodal model with 18B active parameters.

1-bit model runs on 102GB combined RAM + VRAM
Up to 1M context for text, images and documents
Low, High and Max reasoning modes
Improved coding, agent and vision performance

Unsloth improvements

Large GGUFs automatically split across GPU and system RAM
See memory estimates before loading
Chats recover after disconnects
Better multi-image vision and MCP image support
Export chats as JSONL
Improved Auto Compaction controls
Fixes for Linux, NVIDIA Wayland, AMD and Windows

To run or train the new models, you can download Unsloth Desktop:

Download for macOS
Download for Windows
Download for Linux

Thanks for the support for Qwen3.8-27B and Unsloth Desktop last week! For our new v0.1.801-beta release, we merged 200+ PRs to introduce many new features, fixes including:

Auto Compaction (Experimental) for longer chats beyond context limits

Remote & LAN Access (Preview) for easy network access without Cloudflare links
Faster Chat - Improved streaming performance, reduced UI lag, and smoother long conversations.
Support for custom llama.cpp builds. Toggles for Cache RAM, Mmap, Mlock, Checkpoints, Spe Decoding KV Cache, Vision On / Off
Unsloth Dynamic v3.0 is released. New Qwen3.8-27B Dynamic v3.0 GGUFs deliver >10% higher top-1 accuracy compared to everyone else. Works with Unsloth.

Auto compaction (Experimental)

Long chats can now exceed context limits by moving older turns into a searchable archive.

Older turns are removed only when needed and remain searchable. Fresh context epochs improve recall without permanently trimming chats. Uses retrieval instead of summarization for better accuracy.

Remote & LAN access (Preview)

Access Unsloth from other devices on your network.

New remote access settings. LAN control, QR codes, and auto-start options. Disabled by default for security.

Chat improvements

Faster long chats and improved threading. Projects organize chats, files, and workspaces. Added prompt queueing, shortcuts, edit_file, and better tool support.

Hardware, inference, API

Support for custom llama.cpp builds and Intel XPU. More inference controls (cache, mmap, mlock, checkpoints, speculative decoding, vision). Improved GPU validation, VRAM handling, and compatibility. Structured outputs in Responses API. Better llama-server recovery.

To run or train Qwen3.8, you can download Unsloth Desktop:

Download for macOS Download for Windows Download for Linux

Qwen3.8-27B and Qwen3.8-2.4T can now be run locally in Unsloth!

Run on 17GB RAM via Unsloth Dynamic GGUFs. You can also fine-tune Qwen3.8-27B in Unsloth. Qwen3.8-27B is by far the strongest model for its size. We also uploaded NVFP4 quants.

Other Unsloth updates include:

• Qwen3.8-27B + extra llama-server arguments allowed + custom VRAM toggle
• External provider has tool calling + tool support + login with Codex
• Fast FP8 10x faster MiniMax-H3 inference (3 minutes vs 30)
• 10% faster inference for GGUFs + Bypass permissions fixed
• Connected API providers can use their own Search or Unsloth Desktop's built-in Search and tools. Tool results are passed back to the model so it can continue multi-step tasks.
• Sign in with a Codex subscription and use Codex tools inside Chat.

To run or train Qwen3.8, you can download Unsloth Desktop:

Download for macOS Download for Windows Download for Linux

Introducing Unsloth Desktop 🦥 - the first desktop app to run and train models locally.

Open-source. Runs on Mac, Windows and Linux

• Supports MLX, diffusion image/video, audio, GGUF
• Connect Claude Code and Codex to local LLMs
• 50% more accurate, self-healing tool calls + sandboxed code exec
• Works for CPU + multiGPU setups - NVIDIA, AMD, Intel, Mac
• Train models 2× faster with 70% less VRAM
• Private web search, deep research, RAG, MCP + exports (NVFP4, GGUF)
• Use Unsloth’s OpenAI-compatible API and cloud models
• Securely deploy LLMs remotely and access anywhere

You can download Unsloth Desktop now:

Download for macOS Download for Windows Download for Linux

Meta has released Muse Glimmer, the first open model from Meta Superintelligence Labs.

Muse Glimmer is a dense, 30B parameter model by Meta, built for local agentic and coding workflows. Released under the Apache 2.0 license you can run Muse Glimmer 30B via Unsloth and Unsloth Dynamic quants for the best performance.

Muse Glimmer 30B can run locally on 20GB RAM/VRAM setups, including Mac and GPUs. You can also train and fine-tune Muse Glimmer 30B using Unsloth on 20GB VRAM.

You can run and fine-tune Muse Glimmer via Unsloth Desktop app:

• Download for macOS
• Download for Windows
• Download for Linux