Key takeaways

• MiniMax M2 is optimized for end-to-end coding and agentic workflows, aiming for near-frontier intelligence with low latency and deployment efficiency.
• It is a Mixture of Experts (MoE) model with 10 billion activated parameters (out of 230 billion total) and open weights.
• The model demonstrates strong capabilities in general reasoning, multi-step task execution, and mathematical problem-solving, achieving 85.108% accuracy on AGIEval English and 88.44% accuracy on MMLU.
• MiniMax M2's limitations include occasional arithmetic errors, struggles with consistency in applying fixes, handling complex code structures, and test environment setup, which is reflected in its 39% accuracy on SWE-bench Lite (SWE-agent).
• Its performance in coding tasks is specifically limited by difficulties with consistent application of fixes, test environment setup, and version-specific code requirements, suggesting restricted real-world usage for complex coding scenarios.

Technical specifications

Context length

197k tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~100000B

Open weights

Yes

Licence

Modified MIT

Regions

China

Vendor

MiniMax

Provider model key

minimax/minimax-m2