Atlas

Models

← All models

MiniMax M3

MiniMax · 2026-06-01 · 427B parameters

MiniMax M3 is MiniMax's 427-billion-parameter natively multimodal MoE model released in June 2026, with about 23 billion active parameters and a 1-million-token context window enabled by a new MiniMax Sparse Attention (MSA) architecture. It processes text, image, and video inputs and is designed for frontier coding and agentic workflows, with open weights available for self-hosting.

Benchmark scores

BenchmarkScore
A-Fantasia74.7
A-Fantasia (Valid Attempts with Retries)74.5
A-Fantasia Backwards Spelling100.0
A-Fantasia Chess52.0
A-Fantasia Chess (Valid Attempts with Retries)51.5
A-Fantasia Cube (Valid Attempts with Retries)72.0
A-Fantasia Cube Rotation72.0
A-Fantasia Spelling (Valid Attempts with Retries)100.0
AA-3D v1.0 Elo740
AA-AnalystAgent (pass@1)44.0
AA-AnalystAgent (pass@5)73.8
AA-AnalystAgent (pass^5)10.0
AA-Briefcase Elo1096
AA-Briefcase v1.1 Elo1089
AA-LCR74.0
AA-LCR v1.183.3
AA-Omniscience Index1.4
AA-Omniscience v1.1 Accuracy44.9
AA-Omniscience v1.1 Hallucination Rate91.4
AA-Omniscience v1.1 Index-5.5
AdvancedIF56.3
AdvancedIF (rubric mean score)88.4
AHK-Eval83.3
AHK-Eval Algorithms66.7
AHK-Eval Data Structures50.0
AHK-Eval Date and Time50.0
AHK-Eval Easy91.7
AHK-Eval Hard50.0
AHK-Eval Hidden Cases74.0
AHK-Eval Mid75.0
Loading Atlas data…