Jarvis limited edition 2026

HybridRoute is a token-efficient routing agent built for the 8 published capability categories: factual knowledge, math reasoning, sentiment classification, summarization, named entity recognition, code debugging, logical reasoning, and code generation. The core idea: local inference costs zero tokens on the leaderboard, so a quantized Qwen2.5-3B model bundled directly in the Docker image handles sentiment, summarization, and named entity recognition, all categories verified correct through repeated real-prompt testing. Categories needing precise multi-step reasoning or syntactic correctness (factual knowledge, math, logic, code debugging, code generation) route to Fireworks, using MiniMax M3 for reasoning-heavy tasks and Kimi K2.7 Code for programming tasks, chosen per category through direct side-by-side accuracy and token-cost comparison rather than assumption. The agent includes a fallback chain if a preferred model fails or returns an empty response, and every remote call is capped with a category-specific token budget tuned from real failure cases encountered during testing, including a bug where models silently truncated answers while reasoning internally despite explicit instructions not to. The full pipeline was validated end-to-end under the actual 4GB RAM / 2 vCPU grading constraints before submission, with repeated clean runs confirming stability.