MiMo-V2-Pro is Xiaomi's flagship foundation model released in March 2026, featuring over 1 trillion total parameters with 42 billion active, designed to serve as the brain of agent systems. It supports a 1M-token context window and is deeply optimized for agentic scenarios with strong tool-call generation, multi-step reasoning, and high schema adherence for structured JSON outputs. On PinchBench and ClawEval agent benchmarks, it achieves globally leading results, with coding ability surpassing Claude Sonnet 4.6 and general agent performance approaching Opus 4.6.
API|Reasoning|Proprietary Model
Knowledge Cutoff
Unknown
Input → Output Format
Context Memory
1.0MIN131KOUT
AI Performance Evaluation
Arena Overall Score
1448
±6As of 2026-04-23
Overall Rank
No.39
12,059 Votes
Arena by Ability
Hard Prompts
1476±7No.33
Expert Knowledge
1507±19No.8
Instruction Following
1449±10No.27
Conversation Memory
1467±13No.29
Creative
1421±14No.38
Coding
1503±11No.29
Math
1449±19No.31
Arena by Occupation
Creative Writing
1437±11No.30
Social Sciences
1479±13No.25
Media
1425±13No.34
Business
1460±13No.24
Healthcare
1479±20No.26
Legal
1459±20No.37
Software
1487±9No.34
Mathematics
1458±23No.31
Source:Arena Intelligence
Overall
AA Intelligence Index
49%↑11%
LiveBench
58%↓2%
Reasoning & Math
GPQA Diamond
87%↑6%
HLE
28%↑11%
LB Reasoning
70%↑10%
LB Math
77%↑3%
LB Data
49%↑0%
Coding
AA Coding Index
41%↑7%
LB Coding
69%↓5%
LB Agentic
30%↓13%
TAU2
95%↑22%
TerminalBench
41%↑10%
SciCode
43%↑2%
Language & Instructions
IFBench
69%↑12%
AA-LCR
61%↓1%
LB Language
69%↓3%
LB IF
43%↓3%
Output Speed
Standard Mode
67tok/s↓15
First Output 45.43s
