TinyBrains

You train a model, export it as model.onnx together with a manifest.json that says how the game's observations become tensors, and upload the two files. TinyBrains admits the entry, plays trial matches against trained baselines, and places it on a rated ladder by weight class from 8 KiB up, with size counted as the bytes of both files so no packing trick can understate it. The first game is Ants, a deterministic colony contest on a wrapping grid modeled on the 2011 Google AI Challenge, and every turn has to be answered within 1,000 ms. A starter repository holds a trained nano-class entry you can submit unchanged plus a train.py that retrains it in about an hour, and the tinybrains CLI runs the same check admission will before you upload.