MicroLLM Lab – Try 7 tiny LLM's in the browser

MicroLLM Lab is a Hacker News-shared web tool that lets users try seven tiny LLMs directly in the browser. Users pick models and run them, with settings such as max new tokens and runtime estimates based on the last observed tokens-per-second rate.

Instead of judging output by writing quality, the lab uses objective checks—regex matches and exact-token comparisons. A 135M-parameter model is allowed to fail, and that failure is treated as the measurement. Results cover speed (tokens per second, sustained decode, suite wall time) and accuracy (pass rate on objective tests) from runs in the browser, and all numbers stay on the local machine.

Charts use the latest suite per model. The tool asks for a name or handle and device/hardware information, and it supports writing a benchmark in JavaScript: the editor is eval()'d in the same origin, and each check runs on the model's decoded text.