one year on
Alibaba open-sources Qwen-Image, a 20B-parameter foundation model that handles Chinese text rendering natively
The release targets a persistent blind spot in AI image generation—accurate text in both alphabetic and logographic scripts—while also claiming SOTA on multiple editing and generation benchmarks.
Alibaba’s Qwen team released Qwen-Image today, a 20-billion-parameter image generation foundation model built on an MMDiT architecture. Qwen-Image is available on Hugging Face, GitHub, and Alibaba’s ModelScope. Qwen-Image’s headline feature is its ability to render complex text within images: the team demonstrated multi-line layouts, paragraph-length captions, and fine-grained details in both English and Chinese, a capability that has long been a weak point for image generation models.
In benchmark evaluations, Qwen-Image claims state-of-the-art results across GenEval, DPG, and OneIG-Bench for generation, as well as GEdit, ImgEdit, and GSO for editing. The model notably outperforms existing systems on Chinese text rendering benchmarks. The team also highlighted precise image editing operations including object addition and removal, style transfer, and character pose adjustment.
The release adds another open model to Alibaba’s growing Qwen lineup.
The record
One year later — open only if you can handle spoilers
The bet paid off. Qwen-Image became one of the most widely used open image models, quickly ported into ComfyUI and served on Hugging Face and NVIDIA's platform, and Alibaba iterated fast—an editing variant, Qwen-Image-Edit, landed August 19, followed by a steady run of updates through late 2025. A leaner 7B Qwen-Image-2.0 arrived in February 2026, topping open image-generation and editing leaderboards. As with open LLMs, the open frontier in image generation increasingly shipped from China.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy