Loading…
A small general-purpose encoder (about 11M parameters in total, int8, Edge TPU compatible) that answers typed questions about a state, either JSON or plain text: yes/no, pick one option, or a score. It scores 4.73 on Decision Index 0.2.1 (self-measured with the public harness, not an official board entry; GLiNER 2.5 small, 73.9M params: 3.82). The public datasets it learned from carry human labels. The homelab and answer-audit data carry soft labels from Gemma 4.
This demo of github.com/SeraphimSerapis/reef runs the model entirely in your browser (LiteRT.js on WebAssembly). Your inputs never leave this page.
Browser int8 kernels round a few values differently from the native runtime: about 95% of labels are identical and probabilities agree within a few percent.
Loading…