MiniMax

MiniMax was founded in 2021 in Shanghai and has become one of China's most prominent AI labs. While early attention focused on consumer products including the Talkie companion app, MiniMax gained international recognition with its M2-series reasoning models released in early 2026.

The flagship M2.5, released February 12, 2026, is a 230-billion-parameter mixture-of-experts model with 10 billion active parameters. It achieves 80.2% on SWE-Bench Verified, a standard coding benchmark, placing it among the top models globally. It supports a 205,000-token context window and is designed around agentic workflows: tasks that require using tools, writing and running code, browsing the web, and completing multi-step processes autonomously.

Pricing is where M2.5 stands apart. At $0.30 per million input tokens and $1.20 per million output tokens, it is 10 to 20 times cheaper than comparable frontier models. MiniMax offers two speed variants: Standard at 50 tokens per second and Lightning at 100 tokens per second. Both have identical benchmark performance. The model supports automatic prompt caching with no manual configuration required.

M2.5 ships as open weights, available for download on Hugging Face. Organizations can self-host the model on their own infrastructure, which is relevant for compliance-sensitive deployments or teams that want to eliminate per-token API costs at scale.

The model includes an architect mode where M2.5 functions as a planning and coordination layer, breaking complex tasks into subtasks and directing other models or tools to execute them. This is particularly effective for large software projects, multi-file code refactors, and complex research workflows.

MiniMax also released M2.7 shortly after M2.5, positioning it as a self-evolving model capable of improving its own performance on repeated tasks.

For developers and teams evaluating agentic AI infrastructure, MiniMax M2.5 represents a serious option: benchmark performance that competes with Claude Opus and GPT-4o at a fraction of the cost, combined with open weights and strong tool-calling capabilities.