Smart-Home GPT-2 · in-browser tool-calling demo

This is a 124M-parameter GPT-2 fine-tuned to read a smart-home command and emit the matching structured tool call. The model and the optional speech recognition run entirely in your browser via WebGPU — no server, no API, nothing leaves your device.

1 Pick an example command (or speak / type your own)
2 Load the model — first run streams ~330 MB from the Hugging Face Hub
3 Hit Generate and watch it produce the JSON tool call

Choose a command

Try a realistic command

Casual, lifelike phrasings — implied rooms, vague values, questions — for validating the assistant on real end-user input. Or edit the prompt in Advanced below.

Load the model

Model

Backend

Precision

idle

Generate the tool call

Advanced — prompt & decoding options

Prompt (SYSTEM + candidate functions + USER command)

The preset picker above rewrites this automatically. Edit it directly to test custom tool lists or phrasings.

Max new tokens

Temperature

Constrained decoding (JSON schema) Typed args (enum / numeric) Wide names (registry ∪ prompt) Retrieval pre-rank (MiniLM, top-K) K

Load a model first ↑

Benchmark & timing details

Run a generation to see prompt tokens, throughput and grammar overhead.