Blog

Research notes from Q Project — how our models actually work, what broke along the way, and what we checked before saying so.

September 2026

Q-Omni: a model with no embedding table

A 53M-parameter model with ternary weights, a frozen 512-bit fingerprint vocabulary shared across text/image/audio, exact-arithmetic circuits, and a 33 MB packed format that runs client-side in a browser tab.

ternary weights frozen vocabulary multimodal webgpu