Chinese AI startup MiniMax has launched MiniMax-M2, a new large language model available under a permissive MIT license for free commercial use. The company states the model is specifically designed for complex “agentic” tasks, where AI acts independently to use software tools, and for programming.
According to a report by VentureBeat, citing the independent benchmarking organization Artificial Analysis, M2 ranks as the top-performing open-weight model on its Intelligence Index. The report notes its performance in agent-related tasks is comparable to leading proprietary systems like GPT-5 and Claude Sonnet 4.5.
Efficient and cost-effective
MiniMax-M2 uses an efficient architecture known as Mixture-of-Experts (MoE). This design allows the model to deliver high performance while requiring less computing power, which can reduce costs for businesses. MiniMax has set its API price at $0.30 per million input tokens and $1.20 per million output tokens. The company claims this makes it significantly cheaper and faster than competing models.
The model is accessible in several ways. Developers can download its complete weights from Hugging Face for local deployment, use it through the company’s API, or interact with it via a web application called MiniMax Agent.
