Unsloth’s v0.1.914-beta release adds five Glass themes, including native glass effects on Mac, plus controls for chat backgrounds, greetings and code fonts. It also adds chat navigation, right-to-left message display and CPU text-model training from the Train page.
For generation, Unsloth says supported Llama and Qwen3 models run faster in its Transformers backend on NVIDIA GPUs using Linux or WSL. In its A100 test, 4-bit Qwen3-8B generated 99.1 tokens per second, up from 25.7; the release notes do not provide an independent test.
Why it matters
People using Unsloth can personalize its chat interface and, if they have no GPU, train text models on a CPU; the release does not specify how fast CPU training will be.
