.png&w=256&q=75)
🤓 Latest Submissions
.png&w=828&q=75)
Token Saver
Token Saver takes the Track 1 premise to its logical end: the cheapest premium API call is the one you never make — but only if you can prove the local answer is good enough first. A zero-cost regex classifier sorts each task into one of eight categories, then routes it to a small quantized model baked into the container: Qwen2.5-3B for general tasks and Qwen2.5-Coder-3B for code and math, loaded one at a time to fit the 4 GB / 2 vCPU grading box. Answers are never trusted as text — they are verified. Math and logic are solved by having the model write a short Python program that is executed in a sandbox; generated and debugged code is compiled and run; named entities are emitted through a grammar that makes malformed JSON impossible; sentiment labels are computed in code from extracted aspects, so the label and its justification can never contradict; summaries are checked against their length and format constraints. Only a task that fails its verification escalates to an allowed Fireworks model, called with reasoning disabled and a tight token cap. The result answers every category locally at near-zero recorded token cost while clearing the accuracy gate, with a watchdog guaranteeing a valid, complete results.json inside the ten-minute limit.
13 Jul 2026
