Unsloth v0.1.905-beta Released with Decision Models and ComfyUI

- 1. At a Glance
- 2. Overview
- 3. Key Changes
- 3.1. Support for Training, Validation, and Serving of Decision Models
- 3.2. Native Loading of ComfyUI Models and Diffusion Speedups
- 3.3. Improved Training Performance for Qwen3.5 and Hybrid Architecture Models
- 3.4. Enhanced Safety via Sandboxed Code Execution Environments
- 3.5. Improvements to Desktop Browser and RAG Features
- 4. Supported Models and Hardware
- 5. How to Update
- 6. Releases Since Our Last Article
- 7. Related Articles
- 8. What to Read Next
- 9. Sources
At a Glance
| Item | Value |
|---|---|
| Repository | unslothai/unsloth |
| Version | v0.1.905-beta |
| Published | 2026-10-09 |
| License | Apache-2.0 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
unslothai/unsloth version v0.1.905-beta has been released. Unsloth is a local UI and toolkit for running and training LLMs and Diffusion models on local hardware, developed in Python and provided under the Apache-2.0 license.
The most significant change in this version is the addition of a feature to convert and train text and vision LLMs into Jev-style decision models. It is reported that this feature dramatically improves decision-making accuracy from the previous 30% to 80%. For users looking to enhance classification or judgment tasks using local LLMs, this is a release well worth considering for an immediate update.
Key Changes
Support for Training, Validation, and Serving of Decision Models
It is now possible to train arbitrary text or vision LLMs as Jev-style decision models using QLoRA. Trained models can be tested directly from the Decision API settings screen, and judgments can be executed alongside confidence scores for all choices. Furthermore, it supports GGUF format exports for Clef and Laya models built on the Qwen3.5 backbone, serving via llama.cpp including image recognition-capable models, and saving and resuming checkpoints for small-footprint adapters and decision heads. This is a feature that impacts users who want to create and deploy small adapters specialized for decision-making.
Native Loading of ComfyUI Models and Diffusion Speedups
Image and video models in ComfyUI format can now be loaded and run directly from Hugging Face. Unsloth automatically recognizes ComfyUI checkpoints and local model folders, allowing ComfyUI text encoders and VAEs to be utilized directly within Unsloth. In addition, support has been added for running ComfyUI NVFP4 and MXFP8 models, alongside expert pairs for Krea-2, HunyuanImage-2.1, and Wan2.2. Furthermore, INT8 ConvRot is now enabled by default for Qwen-Image-2.1, balancing accelerated generation on supported NVIDIA GPUs with the preservation of image details across all precisions.
Improved Training Performance for Qwen3.5 and Hybrid Architecture Models
QLoRA training processes have been significantly accelerated. On A100 and RTX PRO 6000 environments, training for Qwen3.5-35B-A3B is up to 4.1x faster, and for Qwen3-30B-A3B up to 3.3x faster. In addition, sample packing has been improved for hybrid models such as gated-delta, Mamba2, and short-convolution. Other additions include the ability to train lists of prompts and completion messages as a single conversation, the ability to retain unique questions per row in vision datasets, and features for exporting chats and including system prompts in training data.
Enhanced Safety via Sandboxed Code Execution Environments
To improve safety when models execute code, OS-specific sandboxing has been introduced. Code execution is isolated using Bwrap on Linux, Seatbelt on macOS, and MXC on Windows. Users can check sandbox operational status from the settings screen and choose protection levels. This functions as a safety measure when running agents or code generation models in local environments.
Improvements to Desktop Browser and RAG Features
In the built-in browser of Unsloth Desktop, a feature has been added to ask the model questions about the currently open web page. Additionally, confirmation is now prompted upon downloading from the browser with the ability to choose a destination folder, and pinned pages in the sidebar can now be reordered via drag-and-drop similarly to chat history. Search functionality now gracefully skips unavailable search results to fall back to Wikipedia, and reference links such as [1] within answers have been made clickable. The RAG menu has been improved to allow direct operations for selecting, pinning, and unloading embedding models.
Supported Models and Hardware
This version substantially expands support for newly compatible model architectures, quantization formats, and hardware environments.
Newly Supported Models and Architectures
- Decision models: Supports Jev-style decision modeling targeting arbitrary text and vision LLMs. Clef and Laya models built on the Qwen3.5 backbone can now be exported to GGUF format, and serving via llama.cpp is supported, including models capable of image recognition.
- ComfyUI format and Diffusion models: Supports direct loading of single
.safetensorsDiT models from Hugging Face repositories. It also supports running expert pairs for Krea-2, HunyuanImage-2.1, and Wan2.2 (including Wan2.2-A14B). Additionally, ComfyUI-format NVFP4 and MXFP8 models, as well as hosted FP8 and INT8 files for Wan2.2 (including loading underlow_vramconditions), are usable. For LTX-2.3, processing is embedded to automatically distinguish between distilled and dev versions from weight information. - Language and Multimodal models: Supports QLoRA training for Qwen3.5-35B-A3B and Qwen3-30B-A3B. Sample packing improvements have been made for model groups with hybrid structures (gated-delta, Mamba2, short-convolution), alongside learning empty thought channels during non-thought turns in Gemma-4 (26B and 31B), and fixing behavior for “Thinking off" and “Preserve thinking" settings in Qwen3.6.
- Audio and Embedding models: Voice settings now support quantization choices for various Transcribe ASR models and GGUF-format dictation models. In the embedding model picker, EmbeddingGemma-2 is now displayed preferentially, and a fallback utilizing llama-server has been added for embedding models that cannot be loaded by sentence-transformers.
Hardware and Accelerator Support
- NVIDIA GPU: On NVIDIA A100 and RTX PRO 6000 environments, QLoRA training speedups of up to 4.1x for Qwen3.5-35B-A3B and up to 3.3x for Qwen3-30B-A3B have been achieved. Fast image generation using INT8 ConvRot for Qwen-Image-2.1 is also available on supported NVIDIA GPUs. Furthermore, installation of
mamba_ssmkernels when combining Triton 3.4 or later in Compute Capability 7.5 (sm75) environments is supported, while unnecessary build processes below sm80 have been cleaned up. - Intel Core Ultra Arc iGPU: Judgment processing has been added so that integrated GPUs (iGPUs) are recognized as XPU devices.
- WSL Environment: Processing to properly detect
nvidia-smion every query execution has been added, improving GPU recognition on the Windows Subsystem for Linux.
How to Update
Please carry out updates according to the following procedures based on your usage method.
Updating Unsloth Desktop (Desktop App)
If you are using Unsloth Desktop, download the latest installer matching your OS from the official release page to apply the update. The provided packages are as follows:
- Windows:
Unsloth-Desktop-Windows.exe(x64) orUnsloth-Desktop-Windows-ARM64.exe(ARM64) - macOS:
Unsloth-Desktop-MacOS.dmg - Linux (Ubuntu/Debian):
Unsloth-Desktop-Ubuntu.deb(x64) orUnsloth-Desktop-Ubuntu-ARM64.deb(ARM64 / Ubuntu 24.04 or later) - Linux (Generic):
Unsloth-Desktop-Linux.AppImage(x64)
Updating Unsloth Studio (Web UI / Development Version)
If you have cloned the repository and operate Unsloth Studio in a local environment, fetch the latest changes via Git and re-run the included installation script to update.
On macOS, Linux, or WSL environments, run the following commands:
cd unsloth && git pull
./install.sh --local
unsloth studio -p 8888
If you have deployed to an isolated environment by setting UNSLOTH_STUDIO_HOME, run the script similarly with the environment variable specified.
cd unsloth && git pull
UNSLOTH_STUDIO_HOME="$PWD/.studio"./install.sh --local
UNSLOTH_STUDIO_HOME="$PWD/.studio" unsloth studio -p 8888
On Windows PowerShell environments, run the following commands:
cd unsloth; git pull
.\install.ps1 --local
unsloth studio -p 8888
If you have configured an isolated environment, run it as follows:
cd unsloth; git pull
$env:UNSLOTH_STUDIO_HOME="$PWD\.studio";.\install.ps1 --local
$env:UNSLOTH_STUDIO_HOME="$PWD\.studio"; unsloth studio -p 8888
Additionally, in this release, the package versions pinned within install.sh and install.ps1 have been updated to unsloth>=2026.10.2.
Releases Since Our Last Article
Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.
| Version | Released | Release notes |
|---|---|---|
| v0.1.904-beta (Train your own Decision model) | 2026-10-07 | GitHub |
Related Articles
- Unsloth v0.1.808-beta Released: Major Performance Boosts
- Unsloth v0.1.918-beta Adds Computer Use, MCP Game Engine Support
- Unsloth v0.1.917-beta Released with Major Performance Boosts
- Unsloth v0.1.916-beta Released: Performance & Multi-GPU Updates
What to Read Next
- Follow this tool → Unsloth overview and release history (89 releases tracked)
- Other quantization, model formats and fine-tuning → ggml / ExLlamaV3
- Engines mentioned in this article → ComfyUI
Sources
- unsloth v0.1.905-beta release notes
- Train your own Decision model with Unsloth Official Guide
- Pull Request #12772: Train a decision model from a plain language model
- Pull Request #12876: Train any text or vision LLM as a Clef decision model from Studio
- Pull Request #12939: Serve decision models through llama.cpp in Studio, and export them to GGUF
- Pull Request #12874: Studio: default Qwen-Image-2.1 int8 to the hosted ConvRot file, with shared rotations
- Pull Request #12877: Studio: load ComfyUI nvfp4 and mxfp8 DiT single files
- Pull Request #12878: Studio: recognise ComfyUI checkpoints by name and header, and list ComfyUI model folders
- Pull Request #12879: Studio: load a single.safetensors DiT from any Hugging Face repo
- Pull Request #12883: Studio: load ComfyUI text encoder and VAE files beside a single-file DiT
- Pull Request #12885: Studio: load ComfyUI Krea-2 and HunyuanImage-2.1 single-file DiTs
- Pull Request #12888: Studio: load Wan2.2 hosted FP8 / INT8 files, and hosted files under low_vram
- Pull Request #12944: install-kernels: install mamba_ssm on sm75 with Triton 3.4+
- Pull Request #12960: Bump install.sh / install.ps1 pins to unsloth>=2026.10.2
- Pull Request #12962: Studio: find nvidia-smi under WSL on every query, count Core Ultra Arc iGPUs as XPU
- Pull Request #13005: Studio: use llama-server for embedding models the installed sentence-transformers cannot load
- Pull Request #13006: Studio: prefer llama-server for EmbeddingGemma in the embedding model picker

