MiniMax teases M3 series, M2.7 shows self-evolution

Shanghai-based AI company MiniMax teased the upcoming MiniMax M3 model series on X, while detailing the M2.x family. The M2 series has 229.9B total parameters but only activates 9.8B per token, balancing low activation with high intelligence. It supports a 192K context window and was trained on 29.2T tokens. The model uses a 62-layer decoder Transformer with 256 fine-grained experts, activating 8 per token. MiniMax also introduced Forge, a reinforcement learning system for agents that supports white-box and black-box integration, reducing long-trajectory training costs. The standout M2.7 version shows early self-evolution: it can debug training failures, read logs, modify scaffolding code, and complete 100 rounds of autonomous iteration. The paper says the system now handles 30-50% of the team's daily iteration workload and improved internal coding scaffold performance by 30%, indicating the model is starting to participate in its own development loop.