Z.AI / DOCUMENT INTELLIGENCE / 2026

GLM-OCR

TOTAL1.33B pkgparameters

ACTIVE≈0.9B coreper token

DEPTH24 + 16repeated blocks

HIDDEN1,536main trunk

CONTEXT128Knative tokens

CHECKPOINT2.65GB1 shards · MIT

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

100K INPUT + 1K OUTPUT · 850.34T arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT100.0KIMAGE / PDF → TEXT

TRANSFORMER STACK24L VISION + 16L DECODERCogViT → 2×2 spatial merge → projector → GLM-0.5B-class decoder

encoder–connector–decoder residual chain

OUTPUT1.0KTOKENS

wraps every repeated block

GQA TEXT DECODER · TOKEN DOWNSAMPLINGCogViT 输出的 2×2 相邻 patch 在 connector 中合并，降低语言 decoder 需要处理的视觉 token 数。

CogViT EncoderCogViT 负责读取字形、布局、表格和公式；语言 decoder 不直接承担原始像素编码。

VISION24LCogViT

DECODER16LGLM

MERGE2 × 2spatial

HIDDEN1,536decoder

CONTEXT128Ktokens

PACKAGE2.65GBsingle shard