Model

MiMo-V2.6-Pro

OPEN

Xiaomi · MiMo family · released Sep 22, 2026

Xiaomi's 2026-09-22 flagship, released with open MIT weights alongside a technical report, training environments and RL code. Same 1.02T / 42B MoE scale as V2.5-Pro but now natively omnimodal (text, image, video, audio); the release centres on a single mixed RL run across coding, agent, visual and cybersecurity tasks plus multi-teacher on-policy distillation (MOPD2). Xiaomi's table shows huge gains over V2.5-Pro (DeepSWE v1.1 19.0 -> 71.9, Terminal Bench 2.1 65.2 -> 89.9) and a CyberGym 94.0, while ExploitBench (47.9) stays far behind the closed frontier (70-78.5). API price unchanged at $0.435/$0.87 per Mtok; an UltraSpeed serving tier is sold separately.

ReasoningCodingVisionFunction callingTool useAgentic

3019.9

Elo · rank #5

Parameters

1020B

Active params

42B (MoE)

Context

1M tokens

Architecture

Sparse MoE, 1.02T total / 42B active; 70 layers (60 SWA / 10 GA), 384 routed experts with 8 active, no shared experts; 681M MiMo ViT + 308M AudioTokenizer + 127M audio patch encoder; 5-layer MTP speculative decoder predicting 7 tokens. Natively omnimodal (unlike text-only V2.5-Pro)

License

MIT

Languages

—

API price (in/out)

$0.435 / $0.87

Modalities

text · vision · audio · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AutomationBenchAgents53.1%#2

best: DeepSeek-V4.1-Flash · 54.8%