How fast does your device run a real AI model? A 135M-parameter language model (SmolLM2-135M-Instruct, int8) runs entirely in your browser and generates a fixed 64 tokens from an identical prompt, single-threaded via WebAssembly. Lower time is better.
Running an AI model used to mean sending your request to a server with a rack of GPUs behind it. WebAssembly and quantized ONNX weights make it possible to hold a real language model — not a toy, an actual instruction-tuned transformer with 135 million parameters and 30 layers — inside a browser tab and run it on whatever CPU the visitor's device happens to have. This benchmark measures exactly that: how fast your own machine can do the arithmetic a language model does when it writes.
The model is SmolLM2-135M-Instruct, quantized to 8-bit integers and run through ONNX Runtime's WebAssembly backend, pinned to a single thread so that every visitor's result depends on CPU speed rather than core count. Every participant is given the same fixed prompt and asked to generate exactly 64 tokens with greedy decoding — no sampling, no randomness, no user-editable prompt to skew the comparison. A short, discarded warm-up generation runs first so the timed run isn't paying for the ONNX session's one-time kernel-selection cost.
The score is the wall-clock time in milliseconds for those 64 tokens, timed with performance.now() from the first generated token to the last — lower is better. Reference points: a recent laptop CPU typically lands in the low hundreds of milliseconds per token if unthrottled cores are scarce, faster with more headroom; a phone is usually several times slower. The generated text itself is shown for interest only and is never evaluated — this measures speed, not the model's writing.