MiniMax has launched the M2.5 model, built on the MoE architecture with 10 billion active parameters out of 229 billion total. The model was trained in hundreds of thousands of complex environments and is capable of autonomous planning without explicit user instructions.

According to benchmarks, M2.5 outperforms GPT-5.2, Gemini 3 Pro, and Claude in web search, agent tasks, and office-related tasks. In coding tests, it surpasses Claude Opus 4.6. The model weights are available on Hugging Face under the MIT license.

A closed version, M2.5-Lightning, delivers 100 tokens per second—twice as fast as top competitors. Continuous operation costs $1 per hour, with four instances running year-round for $10,000. Developers can access the API and subscription plans on the MiniMax platform.

For more information, visit MiniMax M2.5 announcement.