Unsloth v0.1.808-beta Released: Major Performance Boosts

September 20, 2026

Unsloth v0.1.808-beta Released with Speedups for Diffusion

At a Glance

Item Value
Repository unslothai/unsloth
Version v0.1.808-beta
Published 2026-09-10
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

Unsloth is a local UI tool for running and training LLMs (Large Language Models) and diffusion models. It supports a diverse range of model architectures, including GGUF, MLX, Qwen3.8, DeepSeek-V4, Gemma 4, and FLUX.

In this version, v0.1.808-beta, substantial performance and reliability improvements have been introduced. Notably, diffusion model execution speed has increased by 1.2x to 1.7x, and performance via Vulkan in AMD environments has improved by 20%. This is a very valuable update for users who prioritize acceleration in diffusion models or specific hardware (such as Strix Halo).

Key Changes

Diffusion Model Acceleration

Across INT8 and FP8 pathways, diffusion processing has become 1.2x to 1.7x faster for all models. For users who frequently run image generation and other diffusion models locally, this is an important improvement that directly reduces generation wait times.

AMD and Windows Environment Optimizations

Strix Halo and Strix Point devices now default to using Vulkan to speed up inference. Additionally, on Linux environments using AMD iGPUs without ROCm, the backend has been changed to use Vulkan instead of the CPU. Users with AMD GPUs can expect smoother inference.

Apple Silicon (MLX) Improvements

On Apple Silicon environments, Gated-delta model training speed has increased by up to 25%. Furthermore, quantized MLX KV caches reduce prompt memory usage by up to 74% compared to before. This is a significant benefit for users running LLMs on Mac environments with limited memory resources.

Improved Inference Performance on Strix Halo

On Strix Halo devices, prompt processing has become 23% faster and generation speed has improved by 8%. It is also reported that when more VRAM is allocated to the iGPU, inference speeds can increase by 3x. Users leveraging the latest AMD APUs may achieve high performance by reviewing their settings.

Package and Binary Size Reduction

The size of the Unsloth Python package has been reduced by over 60%. In addition, over 250 bug fixes have been implemented, improving overall operational stability and the efficiency of installation and updates.

PyTorch Update

The default PyTorch version has been updated from 2.10 to 2.11. This optimizes operation in the latest library environments. Improvements have also been made to ensure that the supported PyTorch version is retained even when reinstalling while preserving an existing environment.

Supported Models and Hardware

General Diffusion Model Acceleration

All diffusion models are now 1.2x to 1.7x faster in INT8 and FP8 pathways. Users running image generation models can benefit from this without needing to re-convert their models.

Optimization for AMD Strix Series

Strix Halo and Strix Point devices are newly optimized. These have been changed to use Vulkan by default, resulting in a 23% increase in prompt processing speed and an 8% increase in generation speed. It is reported that inference becomes 3x faster particularly on Strix Halo when allocating more VRAM to the iGPU.

Apple Silicon (MLX) Improvements

Training of Gated-delta models on Apple Silicon is now up to 25% faster. Additionally, quantized MLX KV caches have been improved to reduce prompt memory usage by up to 74%, enabling inference with lower memory consumption.

How to Get It

For developers or those installing from the latest source, follow these steps.

For macOS, Linux, and WSL

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

For Windows PowerShell

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

To update an existing installation, run git pull inside the repository, then re-run the installation script for your respective OS.

Related Articles

Sources

Update History

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

Engines and ToolsMLX,Strix Halo,Unsloth,Verified,Vulkan,拡散モデル