Qwen-Image

Qwen-Image is Alibaba Qwen team's 20-billion-parameter open-weights text-to-image foundation model,

released August 4, 2025. It pairs a 20B MMDiT (Multimodal Diffusion Transformer) with Qwen2.5-VL as the text

encoder, and its headline strength is rendering text inside images — English, Chinese, Japanese, and Korean,

including multi-line paragraph-level layouts — plus native image editing (style transfer, object insertion /

removal, text edits).

Provider-reported benchmarks (Qwen technical report, own evaluation): GenEval 0.91,

DPG 88.32 (highest overall in their table), LongText-Bench ZH 0.946 / EN 0.943, OneIG-Bench ZH 0.548 / EN 0.539,

ImgEdit 4.27 — in most cases matching or beating GPT Image 1 [High] and Seedream 3.0 in the same table.

Independent

corroboration: at launch it ranked 3rd overall — and top open-source model — on the public AI Arena text-to-image

leaderboard (10,000+ human pairwise comparisons), per VentureBeat.

Released under Apache 2.0 (commercial use

allowed). Note: this entry covers the original v1.0 release; later Qwen-Image 2.x releases use the non-commercial

Qwen Research License instead.

Provider-reported (Qwen technical report): GenEval 0.91; DPG 88.32 (highest overall in their comparison table); LongText-Bench ZH 0.946 / EN 0.943; OneIG-Bench ZH 0.548 / EN 0.539; ImgEdit 4.27; GEdit-EN 7.56; ChineseWord (ZH) 0.583 — in most cases matching or beating GPT Image 1 [High], Seedream 3.0, and FLUX.1 [dev] in the same table.

Independent: #3 overall and top open-source model on the AI Arena text-to-image leaderboard (10,000+ human pairwise votes), reported by VentureBeat at launch.