Atlas

Models

← All models

MiMo-V2.5-Pro

Xiaomi · 2026-04-27 · 1T parameters

MiMo-V2.5-Pro is Xiaomi's most capable open-weight MiMo model at release, a sparse Mixture-of-Experts text model with approximately 1.02T total parameters and 42B active per token, 60 sliding-window and 10 global-attention layers, and a 1,048,576-token context window. It was pretrained on a reported 27T tokens using FP8 mixed precision and incorporates three Multi-Token Prediction modules, then post-trained with supervised fine-tuning, agentic reinforcement learning, and multi-teacher on-policy distillation for demanding agentic, software-engineering, and long-horizon tasks.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo879
AA-LCR73.3
AA-Omniscience Index3.6
ALE-Bench1254
APEX-Agents-AA2.4
Arena Code WebDev — Brand & Marketing1461
Arena Code WebDev — Consumer Product1462
Arena Code WebDev — Content Creation Tools1472
Arena Code WebDev — Data & Analytics1464
Arena Code WebDev — Gaming1495
Arena Code WebDev — HTML1486
Arena Code WebDev — React1463
Arena Code WebDev — Reference-Based Design1472
Arena Code WebDev — Simulations1487
Arena Text — Business, Management & Financial Ops (No Style Control)1466
Arena Text — Chinese (No Style Control)1502
Arena Text — Creative Writing (No Style Control)1434
Arena Text — English (No Style Control)1478
Arena Text — Entertainment, Sports & Media (No Style Control)1439
Arena Text — Exclude Ties (No Style Control)1463
Arena Text — Exclude Ties (Style Controlled)1470
Arena Text — Expert (No Style Control)1502
Arena Text — French (No Style Control)1488
Arena Text — French (Style Controlled)1492
Arena Text — German (No Style Control)1449
Arena Text — Hard Prompts (No Style Control)1485
Arena Text — Hard Prompts English (No Style Control)1497
Arena Text — Instruction Following (No Style Control)1475
Arena Text — Japanese (No Style Control)1414
Arena Text — Korean (No Style Control)1429
Loading Atlas data…