Atlas

Models

← All models

MiniMax M2.5

MiniMax · 2026-02-12 · 228.7B parameters

MiniMax M2.5 is MiniMax's approximately 230-billion-parameter MoE reasoning model with about 10 billion active parameters, released in February 2026 and trained extensively with reinforcement learning for coding, search, tool use, and office workflows. Its downloadable weights are governed by the separate MiniMax Model License in LICENSE-MODEL.

Benchmark scores

BenchmarkScore
AA-LCR66.0
AA-Omniscience Index-39.7
AIME 2024 (avg@8)89.2
AIME 2024-202588.8
AIME 2025 (avg@8)88.3
ALE-Bench1033
APEX-Agents6.2
ARC-AGI-163.7
ARC-AGI-24.9
Arena Code WebDev — Brand & Marketing1384
Arena Code WebDev — Consumer Product1367
Arena Code WebDev — Content Creation Tools1435
Arena Code WebDev — Data & Analytics1387
Arena Code WebDev — Gaming1384
Arena Code WebDev — HTML1415
Arena Code WebDev — React1375
Arena Code WebDev — Reference-Based Design1369
Arena Code WebDev — Simulations1436
Arena Text — Business, Management & Financial Ops (No Style Control)1355
Arena Text — Chinese (No Style Control)1390
Arena Text — Creative Writing (No Style Control)1331
Arena Text — English (No Style Control)1383
Arena Text — Entertainment, Sports & Media (No Style Control)1328
Arena Text — Exclude Ties (No Style Control)1325
Arena Text — Exclude Ties (Style Controlled)1371
Arena Text — Expert (No Style Control)1374
Arena Text — French (No Style Control)1358
Arena Text — French (Style Controlled)1398
Arena Text — German (No Style Control)1360
Arena Text — Hard Prompts (No Style Control)1369
Loading Atlas data…