Hybrid Token-Efficient Router
An AI agent that routes every prompt to the cheapest capable model. Local-first strategy: zero Fireworks tokens for simple tasks, cheapest Fireworks model for complex ones.
Local-First Routing
Simple tasks run on a local Qwen 2.5 model inside the container. Zero Fireworks tokens = best possible score.
Cheapest Fireworks Model
Complex tasks route to the most cost-effective Fireworks AI model that meets the accuracy threshold.
Gemma 2 Integration
Gemma 2 9B and 2B models via Fireworks AI for the Best Use of Gemma Models bonus challenge.
8 Task Categories
Factual knowledge, math, sentiment, summarization, NER, code debugging, logical reasoning, code generation.
Scoring Strategy
Judging is based on token count and output accuracy. Local inference uses zero Fireworks tokens while counting fully toward accuracy. The router automatically classifies each task and routes to the optimal provider.