Unsloth v0.1.906-beta Released with Browser Use and Text Diffusion

Unsloth v0.1.906-beta Released with Browser Use and Text Diffusion

At a Glance

Item Value
Repository unslothai/unsloth
Version v0.1.906-beta
Published 2026-10-12
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 v0.1.906-beta of unsloth, a tool for running and training large language models (LLMs) and diffusion models locally, has been released. Unsloth is open-source software written in Python (Apache-2.0 license) that supports running GGUF and MLX format models as well as efficient training of various models.

The most significant change in this version is the addition of the “Browser use" feature, which allows models to directly operate a web browser. By enabling it from the chat menu, local models can autonomously browse web pages, click, input text, scroll, and more. This is a notable update for users who want to delegate web research and browser operations to their models.

Key Changes

“Browser use" for Autonomous Browser Operations

By enabling Browser use from the chat “+" menu, users can now ask models to research or interact with websites. The model operates the browser in a dedicated tab opening next to the chat, executing page loading, clicking, form input, option selection, and scrolling. This feature works with local models and API provider models, and also supports text-only small models. As a safety measure, the specification requires user confirmation when pressing buttons involving logins, form submissions, purchases, deletions, and transmissions, and the design ensures that models do not automatically input passwords, card numbers, or one-time codes. Access permissions for usable sites and download history can be configured from “Settings > Browser".

Support for “Qwen-Image-2.1-Turbo" for 8-Step Image Generation

In image generation features, “Qwen-Image-2.1-Turbo", the official accelerated version of Qwen-Image-2.1, is now supported. It operates in just 8 steps, reduced from traditional generation steps, achieving a 2.5x speedup. It can be run using the 4.2 GB Q4 GGUF model from the “Images" page. Additionally, the update adds support for exporting image LoRA files trained within Unsloth and loading them from external folders, alongside improvements to reuse existing text encoders within the ComfyUI folder to prevent duplicate downloads.

Support for Text Diffusion LLM Training and LoRA+

Model training features have been significantly expanded. The newly established DiffusionTrainer enables the training of text-diffusion LLMs such as DiffusionGemma, LLaDA, LLaDA2, Dream, Nemotron, and SDAR. It also supports LoRA+ training with loraplus_lr_ratio configured, as well as LoRA fine-tuning for GPTQ quantized models. Furthermore, GRPO training via TRL 1.15 now supports training on video prompts and tool calls, and AsyncGRPOTrainer, which balances generation and training on a single GPU, has been introduced. For tool-calling dataset training, multi-turn agent history training retaining tool definitions per data row is possible.

Speedups for VRAM-Exceeding MoE Models and Speculative Decoding for Mac

Execution performance has been enhanced in both inference and training. On NVIDIA and AMD GPU environments, inference speed has improved by up to 3x when running MoE (Mixture of Experts) GGUF models of sizes exceeding the installed VRAM capacity. For Mac (MLX backend) environments, speculative decoding has been introduced, achieving up to 2.8x response speeds while maintaining identical outputs. Draft depth is set automatically, supporting drafters such as MTP, DFlash, DSpark, and EAGLE-3. On the training side, in environments where flash-attn and xformers are not installed (NVIDIA, PyTorch 2.12 or higher), training using sample packing has been made up to 8x faster.

Rendering Stabilization for Long Chats and Bug Fixes

Various fixes related to user interface and operational stability were implemented. Optimizations were made so that text input and scrolling do not become sluggish during long chats or when hundreds of history items exist in the sidebar. In addition, it includes a wide range of behavior improvements such as support for displaying Claude and Gemini thinking processes, context bar display when connected to Ollama, disabling torch.compile on older GPUs prior to Volta, and fixes for rollout crashes and padding when using TRL PPOTrainer.

Supported Models and Hardware

This version adds architecture support and hardware optimizations across a wide range, including image generation models, text diffusion models, and OCR models.

Newly Supported Models

  • Qwen-Image-2.1-Turbo: Supported the official 8-step accelerated version. You can directly load the 4.2 GB Q4 GGUF model from the “Images" page to perform image generation.
  • Text Diffusion LLMs: Enabled training of DiffusionGemma, LLaDA, LLaDA2, Dream, Nemotron, and SDAR through the implementation of DiffusionTrainer.
  • DeepSeek-OCR / DeepSeek-OCR-2: Newly supported as execution models.
  • Qwen2.5-Omni: Supported loading checkpoints specifying Qwen2_5OmniModel.
  • Audio and Multimodal Models: Whisper and SeamlessM4T can now be loaded without going through auto_model, and support for preserving Dia’s audio padding tokens has been added.
  • Decision Models: Supported training of Laya and Clef decision models in Apple Silicon environments. Clef allows data row training and inference using images.
  • SDXL: Supported loading community-created single-file SDXL GGUF models on the “Images" page.

Hardware and Backend Expansions

  • AMD GPU (ROCm): The vLLM engine can now be used optionally as an inference backend in Linux environments and Windows WSL environments. Additionally, fixes were applied to prevent crashes during bitsandbytes training by applying PyTorch for ROCm 6.4 in ROCm 6.0–6.3 environments, avoiding out-of-memory (OOM) issues with VAE tiles in Qwen-Image-2.1, and supporting OpenBLAS thread limits on Windows ROCm.
  • Apple Silicon (Mac / MLX): Supported speculative decoding with the MLX inference engine, allowing MTP, DFlash, DSpark, and EAGLE-3 to be used as drafters. Furthermore, it includes optimizations such as int8 ConvRot processing in the Qwen-Image-2.1 text encoder and freeing the text encoder during denoising processes in unified memory environments to save memory.
  • NVIDIA GPU: Inference processing when running MoE GGUF models of sizes exceeding VRAM has been accelerated by up to 3x. In addition, torch.compile is controlled to automatically turn off on older generation GPUs prior to Volta, and a change was made to fall back to the Vulkan-bundled version of llama.cpp instead of the CPU if the driver is below CUDA 12.4 on Windows.

How to Get It

If you are using a development or local environment built from source code (Unsloth Studio), update the repository and re-run the installer script.

If you are using macOS, Linux, or WSL (such as Ubuntu), run the following commands in your terminal:

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

If you are using a Windows environment (PowerShell), run the following commands to update:

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

If you are using the desktop application (Unsloth Desktop), please obtain the latest installer from the official distribution page to update.

Related Articles

What to Read Next

Sources