Unsloth v0.1.903-beta Released with Built-In Browser and More

At a Glance
| Item | Value |
|---|---|
| Repository | unslothai/unsloth |
| Version | v0.1.903-beta |
| Published | 2026-10-07 |
| 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 “v0.1.903-beta" has been released for “unslothai/unsloth", a UI tool designed for running and fine-tuning LLMs and diffusion models locally. The project is primarily developed in Python and is licensed under Apache-2.0.
The most significant change in this update is the addition of a built-in browser feature integrated right next to the chat interface. This makes previewing HTML generated by models and checking files much smoother, greatly improving development efficiency.
Key Changes
-
Built-in Browser Feature Placed Next to Chat
A browser panel has been added right next to the chat interface. Files such as PDF, Office, HTML, Markdown, code, and text can be opened as tabs. HTML and code created by models are also displayed as tabs with previews and consoles, replacing the previous Canvas feature. It is also possible to highlight part of a page to ask the model questions, or right-click a tab to fork it into the chat. -
Support for EmbeddingGemma 2 Models
Added support for Google DeepMind’s new multimodal embedding model “EmbeddingGemma 2". It features 740M parameters (270M for text-only), optional vision and audio encoders, and supports over 100 languages with an 8K context window. You can import GGUF or safetensors to run locally, and fine-tuning with 4-bit QLoRA or BF16 LoRA is available using FastSentenceTransformer. -
Audio Features for Voice Cloning and Editing
A new audio page has been added, enabling voice cloning from short samples and saving, word editing within recordings, music vocal and drum separation, and voice conversion. It also supports music and sound effect creation, transcription with timestamps and speaker info, and subtitle exporting, alongside over 70 GGUF models. -
Experimental Support for Simultaneous Multi-Model Loading and Fast Inference Engines
A feature to keep multiple models loaded simultaneously has been added (disabled by default). The model specified for each request responds accordingly. Additionally, vLLM and SGLang engines are supported as experimental options for Linux (with NVIDIA GPUs). -
Memory-Efficient Training via Layer Offload
A “Layer offload" feature has been added for LoRA training, which offloads frozen layers to system RAM (compatible with NVIDIA and AMD graphics cards). This makes it possible to train a Llama-3.1-8B 4-bit model with VRAM usage kept down to 7 GiB. -
Speedup for Padding-Free Learning Without Flash Attention
Padding-free learning without flash attention (flash-attn), primarily used in Colab environments, has been accelerated. For instance, in Qwen3-0.6B training, VRAM usage is reduced from 5.45 GB to 1.87 GB, and step speed becomes 5x to 8x faster. -
Enhanced Sandbox Feature for Improved Security
The sandbox feature, which confines code executed by models using Python or terminal tools to a restricted environment, has been enhanced. In addition to OS sandboxes on Linux and macOS, a Windows 11 preview version has been added, and it now also runs in Colab and Docker environments.
Supported Models and Hardware
Newly Supported Models and Architectures
-
EmbeddingGemma 2
Added support for “EmbeddingGemma 2", a new multimodal embedding model developed by Google DeepMind. You can choose between the full 740M parameter model or the 270M text-only model, with optional vision and audio encoders available. It supports over 100 languages and an 8K context window, allowing GGUF and safetensors formats to be imported into Unsloth for local execution. Fine-tuning with 4-bit QLoRA or BF16 LoRA using FastSentenceTransformer is also supported. -
Audio-Related Models
Over 70 GGUF models for tasks such as voice cloning, editing, separation, conversion, music creation, and transcription are now available on the newly established audio page. These are executed via “audio.cpp", which has been newly added as a native engine. -
Diffusion and Other Models
Default ComfyUI settings are now applied to models such as FLUX.1, Qwen-Image, Z-Image, Ideogram 4, Wan2.2, and HunyuanVideo-1.5. In addition, FLUX.1 and Qwen-Image can now be loaded in environments with low host RAM or on fp16-only GPUs. Furthermore, a mechanism has been introduced where GGUF rendering for MiniMax-H3 is maintained on a resident sd-server and released during idle times.
Hardware-Specific Optimizations and Support Status
- AMD Environment
Fine-tuned models can now be exported for Ryzen AI NPU. Additionally, the initial image generation process for Qwen-Image-2.1 has been significantly accelerated on newer Radeon graphics cards. -
Apple Silicon (MLX Environment)
Response speed for the Qwen3.6-35B-A3B model has improved by about 10%, and processing start times for long prompts are up to 10% faster. Furthermore, “Int8 Prefill" has been added as an experimental, optional feature for Macs equipped with M5 chips. Enabling this speeds up long prompt loading times by up to 1.5x at the cost of a slight decrease in answer accuracy. -
NVIDIA Environment
QLoRA training for the gpt-oss-20b model is now 2x to 3x faster. Image generation using Z-Image-Turbo is also up to 20x faster on older generation GPUs. -
Windows 11 (WSL2 Environment)
vLLM and SGLang can now be run within private distributions inside WSL2 (Windows Subsystem for Linux 2) on Windows 11. This makes multi-GPU inference, quantization, and vision features selectable in Windows environments as well.
How to Update
To update Unsloth to the latest version, follow the steps below according to your environment (OS). If you are using developer versions, nightly builds, or experimental installations, navigate to your repository directory, fetch the latest source code, and re-run the installation script.
Updating on macOS, Linux, and WSL Environments
Open your terminal and run the following commands in order:
cd unsloth && git pull
./install.sh --local
unsloth studio -p 8888
Updating on Windows (PowerShell) Environment
Open PowerShell and run the following commands:
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.
| Version | Released | Release notes |
|---|---|---|
| v0.1.902-beta (Command Palette + Desktop UI/UX) | 2026-10-01 | 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 (89 releases tracked)
- Other quantization, model formats and fine-tuning → ggml / ExLlamaV3

