MiniMax outlined development plans for its MiniMax M3.1 and MiniMax M3Pro models at an Aug. 26 interim earnings call, identified in earlier coverage as its 2026 mid-year earnings call. M3Pro's parameter scale is expected to rise to about 3 trillion, with broader reinforcement learning and long-horizon task training aimed at improving generalization and raising the model's intelligence ceiling. Founder and CEO Yan Junjie said the next phase will coordinate gains in model intelligence, inference efficiency and infrastructure, while the company accelerates releases across its M-series and H-series. MiniMax is also adapting M3 and H3 to domestic chips, with a large-scale domestic computing cluster expected to begin handling real production traffic soon.