{"id":2492,"date":"2026-09-23T01:20:01","date_gmt":"2026-09-22T16:20:01","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/09\/23\/unsloth-qwen-image-2-1-agent-skills-update\/"},"modified":"2026-09-23T07:39:50","modified_gmt":"2026-09-22T22:39:50","slug":"unsloth-qwen-image-2-1-agent-skills-update","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/23\/unsloth-qwen-image-2-1-agent-skills-update\/","title":{"rendered":"Unsloth Update: Qwen-Image-2.1 Support and Agent Skills Added"},"content":{"rendered":"

<\/p>\n

At a Glance<\/h2>\n
\n\n\n\n\n\n\n\n\n\n
Item<\/th>\nValue<\/th>\n<\/tr>\n<\/thead>\n
Repository<\/td>\nunslothai\/unsloth<\/a><\/td>\n<\/tr>\n
Version<\/td>\nv0.1.812-beta<\/a><\/td>\n<\/tr>\n
Published<\/td>\n2026-09-23<\/td>\n<\/tr>\n
License<\/td>\nApache-2.0<\/td>\n<\/tr>\n
Source type<\/td>\nPrimary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

Values determined by this site’s code at collection time. Dates are JST.<\/em><\/p>\n

<\/p>\n

Overview<\/h2>\n

Unsloth is a UI tool for running and training LLMs (Large Language Models) and diffusion models in local environments. It supports a wide range of models, including GGUF, MLX, Qwen3.8, DeepSeek-V4, Gemma 4, and FLUX.<\/p>\n

In this release, in addition to adding support for Qwen-Image-2.1, the “Agent Skills” feature has been introduced to guide models in executing specific tasks. Furthermore, the speed of long reasoning blocks in inference has doubled from the conventional 30 FPS to 60 FPS, making this a major update for users who prioritize inference performance.<\/p>\n

Key Changes<\/h2>\n

Addition of Agent Skills<\/h3>\n

Custom skills can now be added to steer models toward specific tasks. It is also possible to reuse existing skills from Claude Code or .agents<\/code> folders. By using @<\/code> in the chat, you can easily select and invoke managed skills.<\/p>\n

Improved Inference Performance<\/h3>\n

Processing speed for long reasoning blocks has been improved. Specifically, it achieves double the speed, going from the conventional 30 FPS to 60 FPS. A smoother experience can be expected for users utilizing models accompanied by a thinking process (Thinking).<\/p>\n

Qwen-Image-2.1 Support<\/h3>\n

Support for Qwen-Image-2.1 has been added, enabling image generation and editing. Additionally, Vision models can now view images returned by MCP tools.<\/p>\n

Improvements to Chat and Project Management<\/h3>\n

Chats in the sidebar can now be dragged to reorder, pinned, or moved into projects. It is also possible to directly edit project names, instructions, and folders from the project page. The Thinking UI\/UX has also been revamped to provide a cleaner, more modern experience.<\/p>\n

Improved Training Reliability<\/h3>\n

Reliability regarding training resumption (resume) and dataset preparation has been improved. Hugging Face datasets now work within recipes, making it possible to resume past training runs, accurately display remaining time after resumption, map dataset columns, and respect selected splits and subsets.<\/p>\n

Linux\/Desktop Environment Improvements<\/h3>\n

Distribution for Linux environments has been significantly improved. In-app updates for Debian installations have been added, along with a native Ubuntu 24.04+ installer for ARM64. Additionally, a new AMD ROCm Docker image including Unsloth Studio and JupyterLab is available.<\/p>\n

Storage and Model Management<\/h3>\n

Chats can now reuse downloaded models, eliminating the need to fetch the same model twice. It is also now possible to check and delete cache occupying disk space from the settings screen.<\/p>\n

Supported Models and Hardware<\/h2>\n

With this update, the supported model architectures and hardware environments have been significantly expanded.<\/p>\n

Model Architectures and Features<\/h3>\n

Qwen-Image-2.1 is newly supported. This enables image generation and image editing. Additionally, regarding Vision models, they can now see images returned by MCP (Model Context Protocol) tools, enabling more advanced multimodal reasoning.<\/p>\n

Hardware and OS Environments<\/h3>\n

Support for Linux environments has been greatly enhanced. An in-app update feature has been added for users utilizing Debian installations. Furthermore, for users utilizing the ARM64 architecture, a native Linux .deb<\/code> installer targeting Ubuntu 24.04 or later is provided.<\/p>\n

For users utilizing AMD GPUs, a new AMD ROCm Docker image bundling Unsloth Studio and JupyterLab has been prepared. In addition, hardware resource management has been optimized, such as improving iGPU (integrated GPU) environments using Vulkan to properly handle unified-memory windows as shared host memory.<\/p>\n

How to Get It<\/h2>\n

The procedures for installing or updating Unsloth from the latest (Nightly) source are as follows:<\/p>\n

Installation and Update on macOS, Linux, and WSL<\/h3>\n

For developer installation (building from the main<\/code> branch), use the following commands:<\/p>\n

git clone https:\/\/github.com\/unslothai\/unsloth\ncd unsloth\n.\/install.sh --local\nunsloth studio -p 8888\n<\/code><\/pre>\n
If you have already cloned the repository, perform the update with the following steps:<\/p>\n
cd unsloth && git pull\n.\/install.sh --local\nunsloth studio -p 8888\n<\/code><\/pre>\n
Installation and Update on Windows (PowerShell)<\/h3>\n
In a Windows environment, perform the installation using PowerShell with the following steps:<\/p>\n
git clone https:\/\/github.com\/unslothai\/unsloth.git\ncd unsloth\nSet-ExecutionPolicy -Scope Process -ExecutionPolicy RemoteSigned\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n
To perform an update, run the following commands:<\/p>\n
cd unsloth; git pull\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n
Advanced Installation Options<\/h3>\n
Depending on specific environments or use cases, you can also install by specifying the following options:<\/p>\n
Installing to an Isolated Location<\/strong> By setting the UNSLOTH_STUDIO_HOME<\/code> environment variable, you can keep the installation complete within a specific directory.<\/p>\n
# Linux\/macOS\/WSL\nUNSLOTH_STUDIO_HOME="$PWD\/.studio".\/install.sh --local\nunsloth studio -p 8888\n\n# Windows PowerShell\n$env:UNSLOTH_STUDIO_HOME="$PWD\\.studio";.\\install.ps1 --local\n$env:UNSLOTH_STUDIO_HOME="$PWD\\.studio"; unsloth studio -p 8888\n<\/code><\/pre>\n
Skipping PyTorch (GGUF-Only Mode)<\/strong><\/p>\n
# Linux\/macOS\/WSL\ncurl -fsSL https:\/\/unsloth.ai\/install.sh | UNSLOTH_NO_TORCH=1 sh\n\n# Windows PowerShell\n$env:UNSLOTH_NO_TORCH=1; irm https:\/\/unsloth.ai\/install.ps1 | iex\n<\/code><\/pre>\n
<\/p>\n
Related Articles<\/h2>\n