Key takeaways

• MiMo-V2.5 is an omnimodal model by Xiaomi, designed for "Pro-level agentic performance" by integrating strong reasoning, rich perception across text, image, audio, and video, and cost-efficiency for agent frameworks.
• It utilizes a Mixture of Experts (MoE) architecture and supports up to 1 million tokens of context, surpassing its predecessor MiMo-V2-Omni in multimodal perception.
• The model demonstrates strong mathematical problem-solving skills with 95.9% accuracy on the MATH-500 benchmark, excelling in algebra, number theory, and prealgebra, but shows minor issues with complex number calculations and interval notation.
• MiMo-V2.5 exhibits strong foundational knowledge in theoretical physics and organic reaction mechanisms (60.48% accuracy on general scientific QA), but struggles with highly specific, nuanced problems, complex multi-step deductions, and numerical precision/unit consistency.
• Key limitations include occasional errors in complex numerical calculations, nuanced problem interpretations, and specific knowledge in domains like industrial polymerization catalysts, coupled with an average duration of 79.9 seconds for General Purpose QA tasks.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text+image+audio+video->text

Parameters

—

Open weights

No

Licence

Proprietary

Regions

China

Vendor

Xiaomi

Provider model key

xiaomi/mimo-v2.5