MiniMax M2

MiniMax's open-weight model optimized for coding and agentic workflows: a Mixture-of-Experts transformer with ~229B total / ~10B active parameters per token, so agent loops get low latency and high concurrency at low cost. An interleaved thinking model — it emits reasoning inside <think>...</think> blocks that must be kept in conversation history across turns (removing them degrades multi-step tool-chain quality, per the official HF model page). 204K-token input / 131K-token output context. Released under the MIT license (HF repo MiniMaxAI/MiniMax-M2, with vLLM/SGLang deployment guides). Powered the free-for-a-limited-time MiniMax Agent product at launch.

Independent (via deeplearning.ai, Oct 2025): #1 among open-weight models on Artificial Analysis' Intelligence Index. Third-party-reported: Terminal-Bench 46.3, xbench-DeepSearch 72, BrowseComp-zh 48.5 (medium.com review — provenance of these numbers not independently verified). Provider-claimed (via TechNode): ~2x inference speed of Anthropic Claude Sonnet 4.5 at 8% of its API cost. Note: a MiniMax M2.7 successor was announced in 2026 but its promised open weights had not shipped as of Apr 2026 (third-party reporting) — track separately.