MiMo-V2-Pro

MiMo-V2-Pro is Xiaomi's flagship foundation model built for real-world agentic workloads. A Mixture-of-Experts model with over 1T total parameters and 42B active, roughly 3x larger than MiMo-V2-Flash. Inherits the Hybrid Attention mechanism with a 7:1 ratio and supports up to 1M-token context. Optimized for coding and agent tasks through post-training scaling across diverse agent scenarios, with native support for complex tool calling and multi-step reasoning.

Benchmark results

Benchmark Score Tags Source
Claw-Eval 61.5% self-reported llm-stats link →
DeepSearchQA 86.7% self-reported llm-stats link →
GDPval-AA 1,426 self-reported llm-stats link →
PinchBench 81.0% self-reported llm-stats link →
SWE-bench Multilingual 71.7% self-reported llm-stats link →
SWE-Bench Verified 78.0% self-reported llm-stats link →
Tau2 Telecom 96.8% self-reported llm-stats link →
Terminal-Bench 2.0 57.1% self-reported llm-stats link →