MiniMax: minimax-m2-her

MiniMax-M2 is a high-efficiency Mixture-of-Experts (MoE) model architected specifically for coding and agentic workflows. While boasting 230B total parameters, it only activates 10B per token, ensuring low latency and high throughput. Its standout feature is "Interleaved Thinking," where the model uses internal reasoning tags (<think>) to plan and self-correct during complex tasks. This makes it exceptionally robust for multi-step tool use, terminal-based coding, and long-horizon planning. It ranks as a top-tier open-weight model, rivaling proprietary giants in agentic benchmarks like VIBE and SWE-bench.