What is minirouter?

MiniRouter is a workspace built for a competition called Gittensor, where participants try to build a small, cheap system that decides which of several language models should answer a given question. Instead of training one giant model to answer everything, the project trains a tiny router that picks the right specialist model and role for each question. The router itself is small on purpose. A frozen 0.6 billion parameter encoder reads a question and turns it into a single vector, and a much smaller head, about 10,000 parameters, uses that vector to decide which model should answer and in what role. This small head is trained using an evolution style method called separable CMA-ES, which breeds many candidate routing heads and keeps the ones that perform best, based only on whether the final answer was right or wrong. The repository wires together three open source models, an automatic grader for math, general knowledge, and coding questions, and tooling for miners (competition participants) to submit and evaluate their own trained routers. It includes training and evaluation code, benchmark configurations, a validator backend that stores submissions in a database, and a public competition website with a leaderboard. A GitHub Actions workflow automatically labels and tracks submission pull requests as they move through review and evaluation. The project reports real benchmark results. Across two tasks, math and general knowledge questions, the trained router scored higher on average than any single model in its pool, because different models are strong at different subjects. The authors also built a diagnostic tool to check how much room a perfect router could still improve on, which showed genuine unused potential on the math task that their current router does not fully capture yet. This is a research and competition codebase written mostly in Python, aimed at people comfortable with machine learning training pipelines, evaluation benchmarks, and running jobs on GPU hardware. It documents its own workflow for contributors, including branch naming rules and where finished model submissions should be placed.