Unsloth v0.1.915-beta Released with Browser Use & Qwen3.8

At a Glance
| Item | Value |
|---|---|
| Repository | unslothai/unsloth |
| Version | v0.1.915-beta |
| Published | 2026-11-05 |
| 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
The latest version of unslothai/unsloth, v0.1.915-beta, has been released. Unsloth is a UI tool for running and training LLMs (Large Language Models) and diffusion models locally. It supports a diverse range of models, including GGUF, MLX, Qwen3.8, DeepSeek-V4.1, Gemma 4, and FLUX.
This update introduces support for Qwen3.8-Flash-Next alongside a “Browser use" feature that enables models to interact directly with files and web pages via an internal browser. It also includes numerous practical improvements such as reduced memory usage, enhanced training efficiency, and a refined UI.
Key Changes
Enhanced Browser and File Operations
By enabling the “Browser use" feature, models can now perform actions such as clicking and typing on HTML or React pages generated within the chat. Additionally, models can directly read and edit documents such as HTML, React, Markdown, code, text files, and even PDFs opened within the browser. Edited content can be reverted to previous versions using the “Back" button.
Furthermore, GGUF models can read up to four web pages at a time without prompting the user for confirmation when tools are executed. This feature brings significant benefits to users who want models to execute complex tasks via a browser or to agent users who require direct document editing.
Reduced Memory Usage and Improved Performance
Startup behavior has been modified so that PyTorch and other ML libraries are no longer loaded automatically upon launch. This significantly reduces idle memory usage. Settings for this startup feature can be changed via Settings > General > Startup performance.
In addition, loading and reloading MLX models have been significantly accelerated, and the time required to reduce the context of loaded models has been shortened to just a few seconds. SDXL rendering has also been accelerated even in severely constrained VRAM environments, including setups with 4 GB of VRAM. These are exceptionally useful improvements for users operating in limited memory resource environments or those who frequently switch models.
Expanded Training Features and Efficiency
When using the latest NVIDIA GPU, specifying float8_training = "fp8_lora" enables faster and lower-memory LoRA training for 16-bit based models. Various training methods such as DPO, ORPO, KTO, GRPO, and CPO have also been made executable in the MLX environment.
This is a critical update for engineers looking to perform fine-tuning efficiently in a local environment.
Chat Experience and UI Improvements
Chat context (history) is now preserved when switching GGUF models. This eliminates the need to reload long histories every time a model is switched.
On the UI side, the design of the currently beta “Glass" theme has been refined. Softer highlights, lighter shadows, and improved frost effects have been applied across menus, sidebars, and the browser overall. A slider has also been added via Settings > Appearance > Composer to adjust the position of the chat box (Composer) between the top and bottom of the screen. These changes improve convenience and visual customizability for users who utilize the chat function daily.
Supported Models and Hardware
This release newly supports the Qwen3.8-Flash-Next model. LiquidAI’s d1 model and Decider 2B (MLX exclusive) have been added to the Decision API, and the latest Kimi models are now displayed in the model picker. On the hardware front, SDXL rendering is supported in environments equipped with 4 GB of VRAM, and low-memory, fast LoRA training via float8_training = "fp8_lora" is available on the latest NVIDIA GPUs.
How to Get It
For developer, nightly, and experimental installations on macOS, Linux, and WSL environments, run the following commands:
git clone https://github.com/unslothai/unsloth
cd unsloth
./install.sh --local
unsloth studio -p 8888
In a Windows PowerShell environment, run the following commands:
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 environment, navigate to the repository directory, run git pull, and execute the installation script again.
Related Articles
- Unsloth v0.1.918-beta Adds Computer Use, MCP Game Engine Support
- Unsloth v0.1.917-beta Released with Major Performance Boosts
- 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 (89 releases tracked)
- Other quantization, model formats and fine-tuning → ggml / ExLlamaV3

