MinimaLabs/KeyLM-75m-Instruct-Demo

KeyLM 75M Instruct — Demo

A chat frontend for KeyLM-75M-Instruct, a 75M-parameter instruction-tuned model.

The model is served by a small inference API. The page wakes it as soon as it opens — the service sleeps when idle, so the first visit after a quiet period takes a moment before it starts answering — and keeps full multi-turn history, trimmed to the model's 2048-token context as the conversation fills.

Every reply reports how it was produced: output tokens, latency, and the decode rate measured inside the server's generation loop. That rate excludes prompt processing, so it describes the speed of the reply itself rather than the round-trip.

Settings — system prompt, temperature, max reply tokens — are stored in your browser only. KeyLM's chat template has no system role, so a system prompt is folded into the first user message rather than sent as a separate turn.

Single file, no dependencies, no build step. Nothing is sent anywhere except the inference API, and no conversation leaves your browser.