Welcome to Apex Chess

Apex Chess is a solo-developed neural network chess engine, built from scratch with no cloning — only a loose inspiration from AlphaZero. The current deployed model, Apex-Light 1.2, is a 44 million parameter dual-head ResNet trained from large-scale evaluation data.

Version 1.2 achieves approximately 46% Top-1 policy accuracy and a Mean Absolute Error (MAE) of 0.10, translating to an estimated playing strength around 2200–2300 Elo. The upcoming 2.0 release is projected to reach 50%+ Top-1 accuracy and 0.08 MAE, with improved inference speed and tighter value calibration.

A full-size Apex model is also under development, expected to surpass 55% Top-1 accuracy and reach an MAE of 0.07 or better — pushing toward top-tier engine benchmarks.