MiniMax-M2

MiniMax-M2 is a lightweight open-source language model from Chinese AGI company MiniMax, built specifically for Agent workflows and coding tasks. Launched in late October 2025, it carries 10 billion active parameters within a 230-billion-parameter Mixture-of-Experts architecture, and was priced at just 8% of Claude Sonnet with double the speed at release . Within a week it ranked among the top three on OpenRouter's global API call volume and hit number one on HuggingFace Trending ; by December 2025, Artificial Analysis scored it fifth globally among large language models and first among open-source entries . The model introduced "Interleaved Thinking" to maintain reasoning chains across multi-turn tool use, and its successor M2.1 — trained internally on MiniMax's own 385-person engineering team before shipping — pushed multilingual coding benchmarks to what the company claimed was state-of-the-art, scoring 88.6 on MiniMax's own VIBE benchmark and approaching Claude Opus 4.5 . By early 2026, the M2 series' daily token consumption had grown more than sixfold from December 2025 levels, and MiniMax had already moved on to the M2.5 generation .