A 48M-parameter tool-calling model, running entirely in this browser tab. No server, no telemetry — a 105KB WebAssembly engine plus 48MB of int8 weights, downloaded once.
Structure is guaranteed by a grammar compiled from the catalog below: the output is always valid JSON, argument keys can only come from the schemas, and undeclared tools cannot be called. Edit the catalog — the grammar adapts instantly, no retraining.
…
The model is consulted at exactly five choice points — refuse or call, which tool, include this optional, what value, stop or continue. Everything else is emitted by the grammar. Built by Nikshep Saravanan; engine is one C file, MIT. Native speed is ~130–350ms per call on an M3; wasm runs the same int8 kernels via SIMD128.