Unsloth released version 0.1.919-beta on 17 November 2026. The update introduces a beta feature called Computer Use, which lets a locally run vision model watch your screen and move the mouse or type in any app. The feature is off by default and asks for permission before each new app or risky action. At the same time, the MiniMax‑H3‑Image model now runs with GGUF files on CPUs, Macs and AMD GPUs, and with FP8 or INT8 weights on NVIDIA cards. The change speeds up a 2048 px image decode from 2.82 seconds to 0.28 seconds on an A100 80 GB GPU.
What else is new?
The release adds a new model picker inside the composer, letting users switch models and set reasoning options from one pane. Project folders have been redesigned with a Files tab that shows Git changes and lets you search within a project. Connections now show live status for 9 new providers, including NVIDIA NIM and Azure OpenAI, and automatically summarize long conversations. LoRA training (a lightweight fine‑tuning method) is supported for several image models such as FLUX.3 Klein and Qwen‑Image‑2.1. Training speed for Gemma 4 12B improves by up to 2.6 times thanks to new attention kernels.
Background
Unsloth’s Computer Use builds on its earlier Browser Use feature, extending local control from the web to the whole desktop. It works on macOS, Windows and Linux when Unsloth runs directly on the machine – not through Docker, WSL or remote notebooks. The feature requires a vision‑capable model (GGUF, safetensors or MLX formats). The documentation lists Qwen‑3.8‑27B as the best‑performing model in their tests, handling tasks like reading receipt photos, filling spreadsheets and creating slide decks.
Why it matters
For people who run AI locally, the update means their models can now automate real desktop tasks—like filling forms or editing images—without leaving the computer. The faster MiniMax‑H3‑Image also makes image generation feel more responsive, especially on high‑end GPUs.