Z.AI / HYBRID IMAGE GENERATION / 2026

GLM-Image

TOTAL9B + 7Bparameters

ACTIVE7B / stepper denoise step

DEPTH40 AR + 30 DiTrepeated blocks

HIDDEN4,096main trunk

CONTEXT1K–2Kmax canvas

CHECKPOINT35.77GB9 shards · MIT

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

1024² · 50 FLOW STEPS · 1.04P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT1024²PIXELS

TRANSFORMER STACKHYBRID GENERATION PIPELINE40L AR generator → 1K–4K visual tokens → 30L DiT × steps

two-stage residual systems · semantic AR once, diffusion trunk repeatedly

OUTPUT1024²DENOISE STEPS

wraps every repeated block

SINGLE-STREAM DIT · COMPACT VISUAL PLAN9B AR generator 先输出约 256 个紧凑语义 token，再扩展为高分辨率视觉 token，随后进入扩散解码。

Glyph EncoderGlyph Encoder 为图中长文本和复杂排版提供字形条件，是 GLM-Image 区别于普通 latent diffusion 的关键分支。

AR9B · 40Lsemantic

DIT7B · 30Lsingle stream

HEADS32 × 128DiT

LATENT16chpatch 2

VQ16,384codes

LICENSEMITweights