main
official config + MSA report · 59 BF16 shards
MINIMAX / SPARSE MULTIMODAL / 2026

MiniMax M3

TOTAL427Bparameters
ACTIVE≈23Bper token
DEPTH60 + ViT32repeated blocks
HIDDEN6,144main trunk
CONTEXT1Mnative tokens
CHECKPOINT854.2GB59 shards · MiniMax Community
FAMILY EVOLUTION
M1LightningM2.7dense GQAMiniMax M3MSA + native vision
INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

MiniMax M3 · Sparse Multimodal Explorer

100K INPUT + 1K OUTPUT · 5.83P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER
INPUT100.0KIMAGE · VIDEO · TEXT
TRANSFORMER STACK60 TEXT LAYERS + 32L VIT3 dense attention · 57 MSA · 3 dense FFN · 57 MoE
pre-norm residual · Gemma norm · native media tokens enter one 1M stream
++
OUTPUT1.0KTOKENS
wraps every repeated block
MINIMAX SPARSE ATTENTION · BLOCK SPARSE INDEXIndexer 在 block 粒度筛选历史位置，把 1M 上下文的主 attention 读集限制在 2,048 个 token 左右。
MiniMax M3 MoE · 427B CAPACITY≈23B active path · total capacity ≠ active compute
+
TEXT60L3 dense + 57 MSA
VISION32Ldynamic res
HIDDEN6,14464Q / 4KV
MSA16 × 128blocks
EXPERTS128top-4 + 1
CONTEXT1Mtokens