MiniMax M2.7 is a next-generation flagship model that builds on M2.5 with a self-evolving training paradigm — autonomously running over 100 rounds of scaffold optimization during training, achieving a 30% performance improvement. It is built for complex agentic workflows including Agent Teams, dynamic tool search, and elaborate productivity tasks. The model scores 56.22% on SWE-Pro (matching GPT-5.3-Codex) and 57.0% on Terminal Bench 2, demonstrating system-level comprehension. Based on a 230B sparse MoE architecture, it offers frontier performance at just $0.30 per million input tokens.
API|Reasoning|Open Modelproprietary
Knowledge Cutoff
Unknown
Input → Output Format
Context Memory
205KIN128KOUT
AI Performance Evaluation
Arena Overall Score
1405
±6As of 2026-05-01
Overall Rank
No.101
12,349 Votes
Arena by Ability
Hard Prompts
1429±7No.94
Expert Knowledge
1435±19No.83
Instruction Following
1399±10No.97
Conversation Memory
1413±13No.92
Creative
1352±15No.124
Coding
1466±10No.77
Math
1408±20No.94
Arena by Occupation
Creative Writing
1373±12No.116
Social Sciences
1409±14No.116
Media
1351±13No.125
Business
1415±12No.81
Healthcare
1419±22No.111
Legal
1406±21No.107
Software
1456±9No.76
Mathematics
1420±22No.82
Source:Arena Intelligence
Overall
AA Intelligence Index
50%↑10%
LiveBench
65%↑4%
Reasoning & Math
GPQA Diamond
87%↑5%
HLE
28%↑11%
LB Reasoning
75%↑6%
LB Math
81%↑6%
LB Data
56%↑3%
Coding
AA Coding Index
42%↑5%
LB Coding
55%↓18%
LB Agentic
50%↑5%
TAU2
85%↑4%
TerminalBench
39%↑5%
SciCode
47%↑5%
Language & Instructions
IFBench
76%↑13%
AA-LCR
69%↑7%
LB Language
67%↓5%
LB IF
61%↑10%
Output Speed
Standard Mode
45tok/s↓32
First Output 56.24s
