main
official Meituan LongCat config + 26 BF16 shard headers
MEITUAN / LONGCAT / 2026

LongCat-Flash-Lite

TOTAL69.07Bparameters
ACTIVE≈3Bper token
DEPTH14repeated blocks
HIDDEN3,072main trunk
CONTEXT327Knative tokens
CHECKPOINT138.19GB26 shards · MIT
FAMILY EVOLUTION
Flash-Lite69B · 14LLongCat-Nextany-to-anyLongCat-2.01.775T · DSA
INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

LongCat-Flash-Lite · Architecture Explorer

100K INPUT + 1K OUTPUT · 2.07P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER
INPUT100.0KTEXT
TRANSFORMER STACK14 VERY-WIDE LAYERSMLA + shortcut-connected MoE in every block
Shortcut-Connected Residual
++
OUTPUT1.0KTOKENS
wraps every repeated block
DENSE MLA · ZERO-COMPUTATION EXPERTSrouter 可把 token 送入不执行 FFN 的 identity expert,动态控制计算深度。
LongCat MoE · 69.07B CAPACITY≈3B active · total capacity ≠ active compute
DEPTH14very shallow
HIDDEN3,072trunk
HEADS32MLA
EXPERTS256top-12
ZERO128identity
CONTEXT327Ktokens