AI company MiniMax released the M3 model featuring a new sparse attention design that significantly reduces computational requirements for inference tasks.