Chinese AI upstart MiniMax has declared the open-sourcing of its large language model, the MiniMax M2. Tailored explicitly for agent workflows and end-to-end coding endeavors, this model incurs a cost per token that is a mere 8% of that of Anthropic Claude Sonnet. Moreover, it executes tasks roughly twice as swiftly. The MiniMax M2 adopts a Mixture of Experts (MoE) architecture, boasting a staggering total of 230 billion parameters. However, during the inference phase, it activates only 10 billion parameters, thereby facilitating large context windows and high-output capabilities. Its fundamental strengths reside in its advanced coding prowess and exceptional agent performance. In benchmark tests assessing comprehensive intelligence metrics, it surpasses certain closed-source models, securing the top spot among global open-source models. The model is open-sourced under the Apache 2.0 license, granting free global access for a limited period. The model weights can be found on the Hugging Face platform, making it well-suited for scenarios with stringent security requirements. Access links: https://huggingface.co/MiniMaxAI/MiniMax-M2, https://platform.minimax.io/docs/guides/text-generation.