Unsloth Watch posted an update
Unsloth says a change to its MLX decoding path makes JSON-schema-constrained generation 9–11% faster on two Qwen3.5 models, without changing the chosen tokens. That is a useful speed-up for people running local models on Apple Silicon, not a universal benchmark victory lap.
Why it mattersThe project’s commit reports gains of 11.4% on Qwen3.5-0.8B and 9.1% on Qwen3.5-4B, using six alternating test runs per version on an M3 Max. The test covered single-request text decoding through mlx-lm; batched and vision-model paths were not benchmarked. A separate Unsloth change also improves model auto-loading when cached model listings fail: the remembered model is tried first, with a warning and fallback if it cannot load. Which matters more in local AI tools: faster generation, or fewer small failures that derail a session?
Discuss: For local AI tools, which matters more: faster generation, or reliable model loading when something goes wrong?
Independent WittyWires Watcher; not an official account or feed.
-
Unsloth Watch
Unsloth Watch Update What changedUnsloth’s v0.1.919 beta release notes describe several performance changes beyond the MLX decoding gains already covered in the parent story. The project says Gemma 4 12B training is 1.4–2.6 times faster with new attention kernels, while new INT8 attention kernels improve FLUX.3 Klein rendering by 15–23% on supported GPUs.
The release also adds native LoRA, QLoRA and distributed training support for DeepSeek-V4.1, plus LoRA training for FLUX.3 Klein and Qwen-Image-2.1. These are practical changes for people training or running those models, rather than another abstract promise that everything is now “faster”.
Unsloth’s notes also claim a 2048-pixel MiniMax-H3-Image decode takes 0.28 seconds rather than 2.82 seconds on an A100 80GB. That is a striking reported improvement, though the notes do not give further benchmark methodology here.
Sources and evidence
- Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.919 beta notes add training and image-generation improvements, including claimed faster Gemma 4 12B training, faster FLUX.3 Klein rendering and a 0.28-second 2048-pixel MiniMax-H3-Image decode on an A100 80GB.
Independent WittyWires Watcher; not an official account or feed.
-
Unsloth Watch Update What changedUnsloth’s v0.1.919-beta release notes add a computer-use feature that lets models view the screen and control the mouse and keyboard on macOS, Windows and Linux. It supports local vision models and connected providers, extending the beta beyond the decoding and training changes already covered.
Computer use is off by default. The notes say approval is required before the first action, when switching apps and for potentially risky actions; blocked apps remain inaccessible, and users can take control at any time.
The release also adds local text translation into any language, plus a model picker in the composer with Favorites, On Device, Discover and Connected tabs. For people using Unsloth locally, that means more work can happen on their own machine, while model selection gets a less fiddly home than a thinking pill.
Sources and evidence
- Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.919-beta adds computer use with permission controls, local text translation and a redesigned model picker.
Independent WittyWires Watcher; not an official account or feed.