Atlas

Models

← All models

MiniMax M2.7

MiniMax · 2026-03-18 · 228.7B parameters

MiniMax M2.7 is MiniMax's approximately 230-billion-parameter MoE language model with about 10 billion active parameters, released in March 2026 as an agentic model designed for complex multi-agent collaboration and autonomous task execution. It is MiniMax's first model to participate deeply in its own training evolution, including building skills and improving learning workflows, and is optimized for coding, productivity workflows, and long-horizon planning tasks.

Benchmark scores

BenchmarkScore
AA-LCR68.7
AA-Omniscience Index0.7
Agents' Last Exam (ALE)5.9
Agents' Last Exam ALE-CLI Pass Rate5.7
Agents' Last Exam ALE-CLI Score14.6
Agents' Last Exam Full-Spectrum Pass Rate3.6
Agents' Last Exam Full-Spectrum Score8.4
Agents' Last Exam Last-Exam Pass Rate0.0
Agents' Last Exam Last-Exam Score1.3
Agents' Last Exam Near-Term Pass Rate10.4
Agents' Last Exam Near-Term Score24.5
Agents' Last Exam Overall Score14.2
Agents' Last Exam Unlicensed Full-Spectrum Pass Rate4.0
Agents' Last Exam Unlicensed Full-Spectrum Score8.2
Agents' Last Exam Unlicensed Last-Exam Pass Rate0.0
Agents' Last Exam Unlicensed Last-Exam Score1.6
Agents' Last Exam Unlicensed Near-Term Pass Rate10.4
Agents' Last Exam Unlicensed Near-Term Score24.5
Agents' Last Exam Unlicensed Overall Pass Rate6.4
Agents' Last Exam Unlicensed Overall Score14.9
AIME 2024 (avg@8)91.7
AIME 2024-202591.0
AIME 2025 (avg@8)90.4
AIME 202687.7
ALE-Bench1021
APEX-Agents-AA10.6
Arena Code WebDev — Brand & Marketing1406
Arena Code WebDev — Consumer Product1407
Arena Code WebDev — Content Creation Tools1402
Arena Code WebDev — Data & Analytics1402
Loading Atlas data…