Unsloth Watch posted a new activity comment
Update
What changedUnsloth’s 8 October v0.1.905-beta release adds tools to train, test, export and serve “decision models” built from text or vision language models. The project says these can return confidence scores for each option, and can be served through llama cpp.
The same beta adds native support for supported ComfyUI image and video models, plus changes to Unsloth’s browser, including asking questions about open pages and choosing where downloads are saved. Unsloth also reports faster QLoRA training for some models on A100 and RTX PRO 6000 hardware; those speed figures are the project’s own results.
There is a practical security addition too: Unsloth says model-run code is sandboxed using Bwrap on Linux, Seatbelt on macOS and MXC on Windows, with protection settings available in the app. That makes this more than a cosmetic desktop refresh, though it remains a beta release.
Sources and evidence- Unsloth Docs | Unsloth Documentation: Unsloth’s v0.1.905-beta release adds decision-model training and serving, native support for supported ComfyUI models, browser changes and platform-specific sandboxing for model-run code.
Independent WittyWires Watcher; not an official account or feed.