The Feedseed sandbox — talk to the model, watch it decide.
on-device function calling · Gemma3-270M · WebGPU · formerly Chuck's
This is Feedseed running entirely in your browser — no server, no account, no cloud round-trip. It's a function-calling model, not a chit-chat model, so its job isn't to make conversation: it's to turn what you say into the right tool call — and, just as importantly, to fire nothing when you're only talking. Load it once (~310 MB), then it works offline forever.
"How far to Raven Ridge?" → distance_to_waypoint. The trail
name is grounded by retrieval, so it's always a real waypoint — never
hallucinated.
"What a beautiful morning" → {"calls":[]}. Firing nothing on
chit-chat is the hard part for a tiny model, and the win here.
"SOS I broke my ankle" → dial_emergency. In the real app the
deterministic layer backs this up — the model is never the only path to help.
Before you load the model — or where there's no WebGPU — the same window answers from the deterministic rule parser. It always responds.
What "compare" shows. This sandbox is built with compare="rules",
so under each answer you'll see what the deterministic rule parser would
have fired. That's the A/B you actually care about: where the 270M model earns
its download over plain regex, and where it doesn't.
The chat window is a self-contained Web Component. Drop it on any page you control — one script tag, one element. Styles are Shadow-DOM sealed, so it can't collide with the host and the host can't collide with it: