Alibaba Debuts Qwen-Image-3.0 as a Closed-Source Model

Alibaba has introduced its latest image generation models, Qwen-Image-3.0-Pro and the standard Qwen-Image-3.0, marking a significant performance leap but shifting to a closed-source distribution. On the Artificial Analysis Image Editing Leaderboard, the Pro version debuted at number six, gaining 83 Elo points over the previous Qwen Pro, while the standard model climbed nearly 100 Elo points to land at number 15. In text-to-image generation, the Pro model reached number nine with a 48-point Elo jump, and the standard version surged 135 Elo points to secure the eleventh spot.

In image editing, Qwen-Image-3.0-Pro ranks just behind proprietary models like MAI-Image-2.5-Pro, which leads with an Elo score of 1271, Reve 2.1 at 1263, GPT Image 2 (high) at 1257, MAI-Image-2.5 at 1256, and GPT Image 1.5 (high) at 1250. For text-to-image, the Pro model sits behind Google's Nano Banana 2 Lite and ahead of ByteDance's Seedream 5.0 Pro. Built for professional workflows, the Pro model supports 4.5k-token prompts, renders text as small as 10px, and natively handles 12 languages and over 20 fonts. It also reproduces fine details like pores and individual hair strands, and simulates web pages, games, and live streams.

Available via Alibaba Cloud Model Studio, the Pro model costs $0.04 per image at 1K resolution and $0.075 at 2K, while the standard tier charges a flat $0.03. Both are free to try in Qwen Studio. However, this release breaks from Qwen-Image 1.0 and 2.0, which shipped with open Apache-2.0 weights on Hugging Face. The new version has no open weights, no license, no technical report, and no model card. For comparison, Qwen-Image-2.0 featured an 8-billion-parameter Qwen3-VL encoder and a 7-billion-parameter diffusion decoder with a DPG-Bench score of 88.32, outperforming FLUX.1's 12-billion-parameter score of 83.84.

For developers, this release presents a trade-off. Those currently routing text-to-image API traffic through GPT Image or Nano Banana now have a cheaper alternative for high-volume tasks requiring long prompts or precise text rendering. However, practitioners who self-hosted Qwen-Image-2.0 to maintain privacy or control costs cannot upgrade to the new version, leaving them reliant on the older open-weights model.