Moka v1 (million.dev)

Moka v1, a compact Go AI model, has been launched with only 110,000 parameters and a mere 108 KB runtime designed for web browsers. This model was distilled from the more extensive KataGo using 20,000 teacher positions, making it significantly smaller—about 157 times less than its predecessor while still achieving a competitive performance level, matching KataGo’s preferred moves 46.3% of the time on held-out positions. Moka plays at an approximate skill level of 10 kyu on a 9×9 board, showcasing its potential to operate effectively within constrained environments. This development is particularly significant for the AI/ML community as it demonstrates a successful effort to deploy advanced AI capabilities within standard web applications, making AI-driven gameplay more accessible. Moka's architecture also implies broader implications for efficiency in ML models, emphasizing the trend of creating lightweight yet capable models that can run seamlessly in-browser, which could pave the way for various applications in gaming and beyond. As a result, Moka not only highlights advances in AI model precision and performance but also promotes the democratization of AI technology, enabling users to engage with intelligent systems without needing extensive hardware resources.