XIAOMI MIMO / EFFICIENT AGENTIC / 2026

MiMo-V2-Flash

TOTAL309.8Bparameters

ACTIVE15Bper token

DEPTH48 + MTP3repeated blocks

HIDDEN4,096main trunk

CONTEXT256Knative tokens

CHECKPOINT313.1GB145 shards · MIT

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

100K INPUT + 1K OUTPUT · 4.93P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT100.0KTEXT

TRANSFORMER STACK48 MAIN LAYERS39 SWA · 9 full · 1 dense FFN + 47 MoE

pre-norm additive residual · attention sink on SWA layers

OUTPUT1.0KTOKENS

wraps every repeated block

SWA ⇄ FULL ATTENTION · ATTENTION SINK滑窗层保留 sink bias，使局部窗口仍能聚合到稳定的全局锚点，缓解纯窗口切断远距离信息。

MiMo MoE · 309.8B CAPACITY15B active path · total capacity ≠ active compute

DEPTH48main layers

HIDDEN4,09664 heads

ATTN5 : 1SWA / full

WINDOW128+ sink

EXPERTS256top-8

CONTEXT256Ktokens