Mahjax: A GPU-Accelerated Mahjong Simulator for Reinforcement Learning in JAX

In three linesMahjax is a fully vectorized Riichi Mahjong simulator in JAX for reinforcement learning on GPU. Achieves 2 million steps/sec on 8 NVIDIA A100s (no-red rules) and 1 million (red rules). Demonstrates training agents from scratch without supervised pre-training.

Summary generated by Claude — human-verified