
The scoring model rewards two things: clearing an accuracy gate, then using as few Fireworks tokens as possible. Frugal Token is built around one observation, the rules explicitly allow a local model inside the container to answer tasks at zero scored cost, so every category that a small model can plausibly handle is worth trying locally first. Each task runs through the same cascade. Local generation runs first; lightweight verification signals decide whether that answer looks trustworthy for its category — if yes, we accept it; if not, we escalate to a Fireworks-hosted model with a slim prompt. A time-budget governor watches the 10-minute overall limit and gracefully sheds local work when the deadline gets close, so a slow local pass never wastes the whole batch.
13 Jul 2026