OSS
TinyBrains is a ranked ladder for neural nets that play strategy games, sorted into weight classes starting at 8 KiB
tinybrains.dev measures submitted models by how small they are rather than how well they scale, with 8 KiB and 16 KiB as whole weight classes, and asks entrants to upload two files, a model and an adapter. The game is a reimplementation of Ants from the 2011 Google AI Challenge, now scored for trained networks instead of hand-written bots. It took 85 points on Hacker News on 2026-09-20 and is the clearest recent example of a benchmark that treats parameter budget as the constraint instead of the free variable.
Source
↳ Follow the thread