Unsloth v0.1.807-beta Released with AMD and Apple Silicon Boosts

September 9, 2026September 20, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

Unsloth v0.1.807-beta has been released. Unsloth is a UI tool for running and training LLMs (Large Language Models) and diffusion models in local environments, supporting a diverse range of models such as GGUF, MLX, Qwen3.8, DeepSeek-V4, Gemma 4, and FLUX.

The most significant changes in this update are major performance improvements in AMD and Apple Silicon environments. In particular, on AMD environments, the inference engine now uses Vulkan by default on specific hardware, aiming to improve inference speed.

Breaking Changes & Deprecations

Because the PyTorch version has been updated to 2.11, rebuilding may be required if your existing environment relies on a specific version. Additionally, users utilizing AMD Strix series or iGPUs without ROCm are advised to verify their operation as the backend switches to Vulkan.

Key Changes

Adoption of Vulkan as Default on AMD Environments

Vulkan is now used by default for inference on Strix Halo and Strix Point. This is reported to improve prefill and decode performance by 20% compared to using ROCm. Specifically, on Strix Halo, prompt processing is accelerated by up to 23% and generation speed by 8%. Furthermore, when utilizing AMD iGPUs without ROCm in Linux environments, Vulkan will now be used instead of the CPU.

Apple Silicon (MLX) Optimizations

Efficiency when using MLX in Apple Silicon environments has been significantly improved. Training speed for Gated-delta models has increased by up to 25%, and prompt memory usage for quantized MLX KV caches has been reduced by up to 74% compared to before. Additionally, fine-tuning using DoRA and more DPO loss types is now possible on Apple Silicon.

Improved Reliability on Windows

llama-server.exe for Windows has been digitally signed. This reduces false positives caused by Smart App Control (SAC). In addition, when model loading is blocked due to code integrity, the system has been improved to clearly explain the reason to the user.

Expansion of Docker Images and Package Lightweighting

New Docker images supporting NVIDIA Turing through Blackwell generations have been released. These include images for training and Studio. Native images for AMD64 and ARM64 are also provided via Docker Hub. In addition, the size of the Unsloth Python package has been reduced by 44%, enabling a more lightweight installation.

Supported Models and Hardware

AMD Hardware

Fast inference via Vulkan is available on AMD Strix Halo and Strix Point. Vulkan can also be used on AMD iGPUs in Linux environments that do not use ROCm.

Apple Silicon

In Apple Silicon environments, training of Gated-delta models and fine-tuning using DoRA and various DPO loss types are possible. Moreover, more multimodal models can now be fine-tuned using text-only datasets.

NVIDIA GPU

New Docker images supporting NVIDIA GPUs from the Turing to Blackwell generations are provided.

How to Get It

Installation and update procedures for developers (macOS, Linux, WSL) are as follows:

git clone https://github.com/unslothai/unsloth
cd unsloth
./install.sh --local
unsloth studio -p 8888

To perform an update, run the following command:

cd unsloth && git pull
./install.sh --local
unsloth studio -p 8888

For Windows (PowerShell), use the following procedure:

git clone https://github.com/unslothai/unsloth.git
cd unsloth
Set-ExecutionPolicy -Scope Process -ExecutionPolicy RemoteSigned
.\install.ps1 --local
unsloth studio -p 8888

The update procedure is as follows:

cd unsloth; git pull
.\install.ps1 --local
unsloth studio -p 8888

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.

Related Articles

• Unsloth v0.1.811-beta Released with AMD & NVIDIA Docker Images
• Unsloth v0.1.808-beta Released with Major Performance Boosts
• Unsloth v0.1.810-beta Released with Multi-User and AMD Support
• Unsloth Releases Windows ARM64 Binary Version

Sources

• unslothai/unsloth v0.1.807-beta

Update History

• 2026-09-19: Rewrote the article from re-collected sources and restored it from draft to published.