MiMo-V2-Omni by Xiaomi

Key takeaways

• MiMo-V2-Omni is an omni-modal model capable of natively processing image, video, and audio inputs within a unified architecture, targeting complex real-world tasks that span multiple modalities. It combines strong multimodal perception with agentic capabilities like visual grounding, multi-step planning, tool use, and code execution.
• The model utilizes a Mixture of Experts (MoE) architecture with a 256K context window, making it suitable for handling extensive input sequences.
• MiMo-V2-Omni demonstrates strong mathematical reasoning, achieving 97.6% accuracy on the MATH-500 benchmark. Its strengths include consistent high accuracy across various mathematical domains and difficulty levels, excellent handling of symbolic manipulation, geometric calculations, and accurate application of theorems, often providing detailed step-by-step explanations.
• The model's weakest performance is in advanced PostgreSQL query generation, with a low 29.33% accuracy on the BIRD-CRITIC benchmark. It struggles with logical errors, recursive CTEs, and specific JSONB operations, indicating significant gaps in understanding complex SQL.
• Minor formatting issues and occasional misinterpretations of problem constraints or diagrams slightly impact its mathematical problem-solving, indicating areas for refinement rather than fundamental reasoning flaws. In SQL generation, it demonstrates foundational knowledge but frequently fails with precise application of advanced constructs.

Technical specifications

Context length

262k tokens

Max output

—

Architecture

MoE

Modalities

text+image->text

Parameters

—

Open weights

No

Licence

Proprietary

Regions

China

Vendor

Xiaomi

Provider model key

xiaomi/mimo-v2-omni