Key takeaways

• MiMo-V2-Pro is Xiaomi's flagship MoE foundation model, featuring over 1 trillion parameters and a 1M context length, specialized for agentic scenarios and orchestrating complex workflows. It is optimized for general agent frameworks like OpenClaw.
• The model demonstrates strong performance in mathematical problem-solving (90.6% accuracy on MATH benchmark), particularly in basic to intermediate algebra, geometry, number theory, and precalculus, consistently providing detailed step-by-step explanations.
• A significant limitation surfaces in NL2SQL tasks (BIRD-CRITIC benchmark), achieving only 28.33% accuracy. It struggles with complex SQL constructs such as advanced window functions, intricate JSONB pathing, and nuanced `ON CONFLICT` clauses.
• While often generating syntactically correct SQL, the model frequently fails on subtle logical details or edge cases in complex prompts in NL2SQL, leading to semantically incorrect results.
• Despite its strengths, the model exhibits weaknesses in nuanced logical deduction, especially in distinguishing necessary/sufficient conditions and interpreting implied false premises, and struggles with specific formatting requirements in mathematical problem-solving.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

—

Open weights

No

Licence

Proprietary

Regions

China

Vendor

Xiaomi

Provider model key

xiaomi/mimo-v2-pro