Key takeaways

• MiMo-V2.5-Pro is Xiaomi's flagship model primarily designed for general agentic capabilities, complex software engineering, and long-horizon tasks, excelling at sustaining complex operations with over a thousand tool calls and improved instruction following.
• The model utilizes an MoE (Mixture of Experts) 'text-to-text' architecture, demonstrating frontier-tier capability, and notably reduces token usage per trajectory compared to other models.
• It demonstrates strong performance in advanced mathematics (93.33% accuracy on AIME 2026), showcasing robust problem-solving, detailed logical reasoning, and handling of complex algebraic, geometric, number theory, and combinatorial problems, with a single failure case in highly intricate combinatorial counting scenarios.
• The model exhibits strong analytical capabilities and step-by-step reasoning across diverse scientific domains, achieving 83.08% accuracy on a General Purpose Question Answering benchmark, performing well in Chemistry and Physics questions.
• A key limitation is its occasional struggle with nuanced interpretation or specialized knowledge in some scientific domains, leading to misinterpretation of specific conditions or overlooking subtle details, particularly in intricate combinatorial counting and certain organic chemistry stereochemistry problems.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~1000B

Open weights

No

Licence

Proprietary

Regions

China

Vendor

Xiaomi

Provider model key

xiaomi/mimo-v2.5-pro