Tag: #webgpu5 results
A compact 290MB 1-bit quantized Bonsai 1.7B model now runs entirely locally in web browsers using WebGPU. The demo is hosted on Hugging Face Spaces by the webml-community. This enables lightweight LLM inference without server dependency.
Cohere launched a speech-to-text model topping the OpenASR leaderboard for English while supporting 14 languages. A WebGPU demo runs it fully locally in the browser using Transformers.js. Source code and demo available on Hugging Face.
Kuma is a new compiler project that converts PyTorch models into portable, self-contained artifacts for browser-based execution via WebGPU. It aims to eliminate Python and server-side dependencies for model deployment.
Qwen released Qwen 3.5 Small multimodal models in sizes 0.8B to 9B for on-device use. A demo runs the 0.8B variant locally in the browser using Transformers.js on WebGPU. Vision encoder is the main bottleneck but execution is impressive.
Google is developing a new safety toggle in Chrome for Android 16 to prevent WebGPU access to phone graphics hardware. The feature was discovered via teardown of Google Play Services 26.10.31 APK. It will integrate into Android 16's Advanced Protection Mode to control Chrome's WebGPU functions.