Unsloth v0.1.811-beta Released with AMD & NVIDIA Docker Images
At a Glance
| Item | Value |
|---|---|
| Repository | unslothai/unsloth |
| Version | v0.1.811-beta |
| Published | 2026-09-19 |
| License | Apache-2.0 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
The latest version “v0.1.811-beta" of “unslothai/unsloth" has been released. Unsloth is a Python-based tool that provides a UI for running and training LLMs and diffusion models in local environments (Apache-2.0 license, 76,373 stars on GitHub).
The biggest changes in this version are the provision of new Docker images supporting both NVIDIA and AMD environments, and the addition of multi-user account features. This improves portability in Docker environments and allows multiple users to share resources while maintaining individual workspaces.
Key Changes
New Docker Images with NVIDIA and AMD Support The main image and Studio setup for CUDA environments have been updated, and Unsloth training patches on GPU hosts have been restored. In addition to the CUDA version, a Docker image for AMD ROCm has been newly added for supported Linux hosts. Studio data is persisted on volumes without pinning application code, improving maintenance such as updates, generated passwords, and port settings. For Docker installation instructions, please refer to the Docker Installation Guide.
Isolated Environments via Multi-User Accounts Accounts can now be created via “Settings > Accounts" using a one-time setup code and individual passwords. If settings match, loaded models can be shared while completely separating and retaining each account’s work in distinct spaces. The existing single-user behavior remains unchanged.
Speedup via MTP Hotfix for Qwen3.8-Flash-Next As a fix from the previous version (v0.1.810-beta), an MTP (Multi-Token Prediction) hotfix for Qwen3.8-Flash-Next has been applied. This doubles the inference speed.
2x Faster INT8/FP8 Image Diffusion Model Inference Inference for image diffusion models at INT8 and FP8 precision is now newly supported. This makes inference processing twice as fast compared to before.
GRPO Improvements and Qwen3.5 Support Improvements related to GRPO (Group Relative Policy Optimization) were made, support for Qwen3.5 models was added, and compatibility with the latest TRL and vLLM was included. Additionally, issues such as a bug when injecting GRPO prompt trim blocks into TRL (PR #11169) have been fixed.
Improved Usability for Chat and Inference Functions Even while loading a local model, users can now edit, reorder, and control queued prompts. Furthermore, GGUF reasoning budgets settings, chat width customization, desktop interface scaling, and a more compact composer have been added.
Hardware Requirements
This version enhances support for the following hardware and model architectures:
- Expanded AMD GPU Support: AMD RDNA1 and RDNA2 architectures are now newly supported.
- Windows ARM64 CUDA Support: Training and inference using CUDA are now supported in Windows ARM64 environments.
- MLX Feature Extensions: In Apple Silicon environments (MLX), video input support, optional MoE (Mixture of Experts) and decoding optimizations, and multimodal chat stability have been improved.
- Ascend NPU Detection: Added support for detecting Ascend NPU devices.
How to Get It
If you are using the Unsloth Studio developer version (main branch), you can update to the latest version with the following steps:
For macOS, Linux, and WSL
cd unsloth && git pull
./install.sh --local
unsloth studio -p 8888
For Windows (PowerShell)
cd unsloth; git pull
.\\install.ps1 --local
unsloth studio -p 8888