Soma is a coordination network that trains a foundation model through competition, not gradient passing. You download the current best weights, improve them with whatever works — RL, fine-tuning, new data — and submit. The architecture is fixed. The objective is singular: next token prediction loss. Winners are rewarded each round. There is no final checkpoint. The bar keeps rising.
The network routes data to models using exact nearest neighbor over embeddings, so participants naturally specialize. A foundation model emerges from many specialized smaller ones, and the system scales horizontally by adding participants. A separate competition for submitting new data keeps the benchmark continuously moving — the model can never overfit to a static evaluation set.
Training is hard and verification is cheap — the same asymmetry behind proof of work. The network verifies your work without repeating it. Your improvements compound with everyone else’s — you just have to beat them first. Winners are ranked on a public leaderboard. The resulting weights are open source.