Alex Dudov
⬤ MiniMax just dropped its M2.1 model, calling it a significant step up from the previous M2 version. The new release scored 64 on the Artificial Analysis Intelligence Index, landing it firmly in the top five open-weights models currently available. What makes M2.1 interesting is its relatively compact architecture—230 billion total parameters with just 10 billion active ones—making it considerably leaner than powerhouses like DeepSeek V3.2 and Kimi K2 Thinking.
⬤ During testing, M2.1 churned through about 90 million output tokens, putting it right alongside DeepSeek V3.2 and roughly 40% more efficient than Kimi K2 Thinking. The pricing structure sits at $0.30 per million input tokens and $1.20 per million output tokens through the first-party API, with the complete Intelligence Index evaluation running $128. MiniMax also notes better hallucination rates and ongoing improvements for agentic tool-use applications.
⬤ Developers can grab M2.1's open weights directly from Hugging Face for self-hosting across various inference frameworks. The model's also accessible through MiniMax AI and Fireworks AI serverless APIs, with additional platform support rolling out soon.
⬤ M2.1's arrival demonstrates how quickly the open-weights AI space is evolving. As performance, deployment options, and efficiency become critical factors, this model's Intelligence Index ranking shows that open-weights systems are increasingly matching their closed-source competitors.
Alex Dudov
