Qwen-Image

Overview

Qwen-Image is a 20B Multimodal Diffusion Transformer (MMDiT) from the Qwen team, paired with a Qwen2.5-VL-7B text encoder. Its standout capability is text rendering inside the image — English and especially Chinese typography, where it leads the open models by a wide margin. Apache-2.0, fully commercial.

Deployment notes

• 80GB VRAM class only (A100 80G / H100). bf16 weights alone are ~55GB: 20B transformer + 7.6B text encoder + VAE, and the runner loads the whole pipeline onto one device without offload.
• ~58GB of weights to download before the first job.
• Native resolution is 1328×1328; this network caps the request at 1024 — the window where diffusion verification is calibrated.
• CFG knob: this pipeline uses true_cfg_scale, not guidance_scale. Supply a negative prompt and it runs at the pipeline default (4.0); leave it empty and guidance is off.