MEITUAN / LONGCAT FRONTIER / 2026

LongCat-2.0

TOTAL1.775Tparameters

ACTIVE≈48Bper token

DEPTH38 + 3repeated blocks

HIDDEN8,192main trunk

CONTEXT262K / 1M trainnative tokens

CHECKPOINT3.551TB194 shards · MIT

FAMILY EVOLUTION

Flash-Lite69B · 14LLongCat-Nextany-to-anyLongCat-2.01.775T · DSA

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

100K INPUT + 1K OUTPUT · 11.60P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT100.0KTEXT

TRANSFORMER STACK38 VERY-WIDE LAYERSMLA + LongCat Sparse Attention + 768-way MoE

Shortcut-Connected Residual

OUTPUT1.0KTOKENS

wraps every repeated block

SPARSE MLA · LONGCAT INDEXERIndexer 扫描历史并为 sparse MLA 提供候选；主 attention 稀疏但 selection 仍有成本。

LongCat MoE · 1.775T HEADER TOTAL≈48B active · total capacity ≠ active compute

DEPTH38 + 3main + MTP

HIDDEN8,192very wide

INDEX32 × 128top-2,048

EXPERTS768top-12

ZERO128identity

PARAMS1.775Theader