MicroLLM Lab brings small language model tests into the browser
The tool runs quantised models with WebGPU and reports browser-run speed and pass rates on objective checks, with results kept on the user’s machine.
MicroLLM Lab offers a browser-based way to chat with and compare quantised small language models using WebGPU. Its description names PetitGPT and SmolLM2 among the models available; the headline advertises seven, but does not list them all.
The tool measures sustained decoding speed and pass rates on objective tests, including regex and exact-token checks. Those results reflect performance on the specified tests, rather than a broad assessment of writing quality.
MicroLLM Lab says measurements come from runs in the user’s browser and remain on the user’s machine. Charts display the latest test suite for each model. The tool may estimate performance using the user’s last recorded tokens-per-second result.
Users can download a performance certificate listing device hardware and peak and sustained tokens-per-second figures. The source provides no independent verification of the results.
The tool also says its editor is evaluated with `eval()` in the page’s origin before checks run against model output.
