Z.AI / LONG-HORIZON AGENTIC / 2026

GLM-5.2

TOTAL753Bparameters

ACTIVE≈40Bper token

DEPTH78+1repeated blocks

HIDDEN6,144main trunk

CONTEXT1Mnative tokens

CHECKPOINT1.51TB282 shards

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

100K INPUT + 1K OUTPUT · 9.96P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT100.0KTOKENS

TRANSFORMER STACK78 MAIN LAYERS3 dense · 75 sparse MoE · 20 full indexers

pre-norm additive residual · deeper 78-layer serial path

OUTPUT1.0KTOKENS

wraps every repeated block

DSA + INDEXSHARE SYSTEMone full index scan feeds a four-layer group

GLM MoE · 753B CAPACITY≈40B active path · total capacity ≠ active compute

DEPTH78 + 1main + MTP

HIDDEN6,144main trunk

INDEX32 × 128share ×4

EXPERTS256routed

ACTIVE8 + 1per token

PRECISIONBF16router FP32