Running an LLM in the browser: WebGPU hit 63 tokens/s, WASM just 1.6

Developers at BEINGS benchmarked in-browser LLM inference via Transformers.js and a Web Worker on October 17, 2026: a 360M-parameter model generated 63 tokens per second on WebGPU versus 1.6 on the WASM fallback. A 1B Llama answered in under half a second, while a 0.5B Qwen produced so much gibberish that half its replies at higher temperature were discarded.

• WebGPU reached 63 tokens/s; WASM fallback managed 1.6 tokens/s on a 360M model
• A 1B model in 4-bit quantization takes about 1.2 GB instead of 5 GB
• 1B Llama answered in under 0.5 seconds in Chrome 155 on an RTX 4070
• Half of 0.5B Qwen replies at higher temperature were discarded as gibberish