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

Turn natural-language smart-home commands into structured JSON tool calls — with a 124M model that runs entirely in your browser. No server, no cloud, nothing leaves your device.

"dim the living room lights to 20%"
   ↓  GPT-2 124M (smart-home fine-tune, in your browser)
{ "name": "dim_light", "arguments": { "room": "living room", "brightness_pct": 20 } }

• 🔒 Private — the model and the optional Whisper speech recognition run 100% in-browser via WebGPU + transformers.js. Weights stream once from the Hub, then cache. No API.
• 🎙️ Voice in 99 languages — speak a command, get the tool call.
• 🧩 Always-valid JSON — constrained decoding keeps the output schema-correct.
• 📚 123 functions, 4096-token context for long tool lists.

Models (streamed on first load, ~330 MB fp16, then cached): lifeart/smart-home-gpt2-v14-ctx4096 (default, 4096-ctx) · lifeart/smart-home-gpt2-v9 (1024-ctx).

Source & docs: https://github.com/lifeart/smart-home-gpt2-tool

This Space hosts only the static front-end build (a few MB); the model weights are fetched from the Hugging Face Hub at runtime.