Unsloth v0.1.920-beta Released with Enhanced AMD NPU Support

At a Glance
| Item | Value |
|---|---|
| Repository | unslothai/unsloth |
| Version | v0.1.920-beta |
| Published | 2026-11-19 |
| 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
Version v0.1.920-beta of “unslothai/unsloth" (Python implementation, Apache-2.0 license), a tool for running and training LLMs and diffusion models in local environments, has been released. The tool supports various formats such as GGUF and MLX, providing desktop applications and Web UIs.
The most significant change in this update is the major enhancement of AMD NPU support. Converted Q4NX models can now be correctly loaded on NPUs, and status indicators based on FastFlowLM information show whether each model actually runs on the NPU. This update is noteworthy for users leveraging AMD NPU or those who want to perform LoRA training on the latest image and video generation models.
Key Changes
Enhanced AMD NPU Support and Management Features
Features for AMD NPU have been significantly revamped. Q4NX models (text models only) converted in the Export page now display as “Converted" in the NPU models list and can be loaded directly onto the NPU. Additionally, a “Convert to Q4NX" tab has been newly added to the Export page. Furthermore, by reading information from the installed FastFlowLM, each NPU model now displays tags such as “Runs on NPU", “Convert, then runs", “Converts, won’t run", and “Not supported", allowing users to check compatibility at a glance.
Functionality has also been added around the CLI. By running the new command unsloth list-npu-models, users can check NPU model sizes, download status, and whether they fit into the memory of their NPU. In unsloth run and unsloth chat, users can now specify model names displayed in the app or FastFlowLM/ repository names in addition to the traditional lemonade: ID.
LoRA Training Support and Model Support for Image and Video Generation Models
LoRA training is now newly supported for LTX-2.5 and MOSS-Image. Trained LoRAs can be reloaded and used in the Video and Images pages. Furthermore, NVIDIA Cosmos3 Nano-FP8 and Cosmos3 Edge-FP8 video models can now be loaded on the Video page without requiring additional NVIDIA packages.
In addition, LoRA training for FLUX.2-klein, Z-Image, and Wan 2.2 5B can now be executed even on GPUs like T4 that do not support bf16. Operating conditions for MiniMax-H3-Image have also been relaxed to allow rendering on 8 GB GPUs.
Performance Improvements in Inference and Training
In environments with sufficient GPU memory to retain activations, LoRA training for diffusion models has accelerated by 20% to 42% per step. Furthermore, in Strix Halo environments on Windows, FLUX.3 Klein INT8 is up to 28% faster per step, and Qwen-Image-2.1 is up to 23% faster. In Strix Halo environments on Linux, FLUX.3 Klein INT8 has also improved to speeds comparable to bf16.
In LLM training, QLoRA training is accelerated by up to 25% by keeping dequantized base weights in free VRAM, such as when running Llama-3.1-8B on an L4. Additionally, stochastic rounding is now adopted by default in bf16 full fine-tuning, improving cases where small update information would otherwise be lost.
Expanded Chat and Desktop Features
Scheduled tasks can now be added from the chat title menu, and a mechanism has been added where responses from periodically executed tasks are replied to directly within the target chat. In addition, the Translate page has been changed to a new layout, enabling language tabs, left-right split view, and document/image upload functions. On Linux environments, computer use features via the desktop portal are now supported in Wayland sessions.
API and Stability Improvements
In the API server function /v1/chat/completions, fixes have been made so that grammar and json_schema specifications are not ignored and are correctly applied. Safety measures have also been introduced to automatically reject GGUF quantization that causes broken output during Q4NX conversion (such as the issue where Qwen3.5 converted from Q4_0 GGUF outputs only slashes) and proceed to the next quantization process.
Compatible Models and Hardware
This version brings significant progress in support and optimization for various models across AMD NPU, AMD GPU, NVIDIA GPU, and Mac (MLX) environments.
AMD NPU and AMD GPU Environments
In AMD NPU environments, models like Qwen3.6-35B-A3B-NPU2 and LFM3 can now be correctly loaded using FastFlowLM model hub names. Compatibility and manageability of models executable on NPUs have been improved.
In AMD GPU environments (such as Strix Halo), generation time for Z-Image in Vulkan on Windows is about 21% faster, and peak memory usage has been significantly reduced. INT8 attention is now also automatically tried for Qwen-Image-2.1 and Wan. Furthermore, an issue where out-of-system-memory occurred with MiniMax-H3-Image when using CPU PyTorch on 16 GB memory Vulkan cards has been fixed. Flash varlen attention has been introduced for packed training on ROCm, making it faster than traditional padding-based training.
NVIDIA Video and Image Models and Accelerators
As NVIDIA video models, loading for Cosmos3-Nano-FP8 and Cosmos3-Edge-FP8 is now possible on the Video page without additional packages. In the LoRA training function for video and image models, new support has been added for LTX-2.5 and MOSS-Image, and LoRA training for FLUX.2-klein, Z-Image, and Wan 2.2 5B can be run even on bf16-unsupported GPUs like the T4.
Additionally, video decoder processing for LTX-2 has been made about 1.7 to 1.8 times faster, and MiniMax-H3-Image can now render even on GPUs with 8 GB VRAM.
Mac (MLX) Environment
In Mac environments, execution setting options equivalent to the GGUF format can now be applied to MLX models. The model picker can now display and load MLX format tags such as Ollama-format gemma4:*-mlx and qwen3.6:*-mlx, and bundled drafter models are also available.
How to Get It
To update Unsloth’s development version and local environment (Unsloth Studio) from an existing environment to the latest state, run the following commands at the root of the repository.
Update procedure for macOS, Linux, and WSL environments:
cd unsloth && git pull
./install.sh --local
unsloth studio -p 8888
Update procedure for Windows (PowerShell) environments:
cd unsloth; git pull
.\install.ps1 --local
unsloth studio -p 8888
Environment variables can be specified to launch or install with specific configurations. For example, to run in GGUF-only mode without using PyTorch, specify UNSLOTH_NO_TORCH=1. Also, to specify Vulkan as the llama.cpp backend, set environment variables and run the installation script as follows:
Linux / macOS (bash):
export UNSLOTH_LLAMA_CPP_BACKEND=vulkan
curl -fsSL https://unsloth.ai/install.sh | sh
Windows (PowerShell):
$env:UNSLOTH_LLAMA_CPP_BACKEND="vulkan"
irm https://unsloth.ai/install.ps1 | iex
If you are using the desktop app version (Unsloth Desktop), please get the latest version from the official download site.
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 |
|---|---|---|
| v0.1.919-beta (Computer Use + Translate + MiniMax-H3-Image + New Model Picker) | 2026-11-17 | GitHub |
Related Articles
- Unsloth v0.1.918-beta Adds Computer Use, MCP Game Engine Support
- Unsloth v0.1.915-beta Released with Browser Use & Qwen3.8
- Unsloth v0.1.913-beta Released: Faster Updates and Image Gen
- Unsloth 0.1.912-beta: Multi-LoRA Batching and Faster Startup
What to Read Next
- Follow this tool → Unsloth overview and release history (91 releases tracked)
- Other quantization, model formats and fine-tuning → ggml / ExLlamaV3

