v0.1.905-beta v0.1.906-beta
unslothai/unslothOct 8, 2026 → Oct 12, 20264 days between releases
What's New
Adds browser automation, text-diffusion training, and faster image generation
New Features
- Browser automation allowing models to control web pages
- Qwen-Image-2.1-Turbo model support for faster image generation
- Text-diffusion LLM training for DiffusionGemma, LLaDA, and Dream
- Faster MoE GGUFs for models exceeding VRAM limits
- Speculative decoding on Mac for 2.8x faster MLX replies
Improvements
- Qwen-Image-2.1-Turbo generates images in 8 steps (2.5x faster)
- Browser use includes tab isolation and security controls
- Up to 3x faster MoE GGUFs for larger models
- Speculative decoding on Mac accelerates inference
- Text-diffusion training capabilities added
Upgrade Notes
Enable browser use via the new + menu in the chat interface. Configure browser access permissions in Settings to manage site restrictions.
The Two Releases
Sandboxing is here!
Oct 8, 2026· 14 days ago
This version introduces sandboxing features for Windows, Mac, and Linux to isolate the code models run. It also focuses on decision models, allowing users to train their own Jev-style decision models with improved accuracy. The release includes native ComfyUI model support and various performance and training improvements.
Full release notesBrowser Use + Qwen-Image-2.1-Turbo
Oct 12, 2026· 10 days ago
This release introduces browser automation capabilities for models, allowing them to interact with web pages directly. It also brings significant performance improvements for image generation and training, specifically with the Qwen-Image-2.1-Turbo model and speculative decoding on Mac. Additionally, it adds support for text-diffusion LLM training and faster MoE GGUFs.
Full release notes