Qwen Image: Alibaba's AI image model — every version from 1.0 to 3.0

What is Qwen Image

Qwen Image is a line of image models from Alibaba’s Qwen team, the same team that ships the Qwen language models. The first Qwen-Image came out on August 4, 2025: 20 billion parameters, an MMDiT architecture and open weights under Apache 2.0. From day one the series stood out for text rendering, starting with Chinese and English.

Within a year the line grew to about ten models, and the version numbers don’t run in order: the closed 3.0 shipped before the open 2.1. Here’s the map:

Version	Date	What it is	Weights
Qwen-Image	Aug 4, 2025	Text-to-image, 20B parameters	Apache 2.0
Qwen-Image-Edit	Aug 18, 2025	Single-image editing	Apache 2.0
Edit-2509	Sep 22, 2025	Multi-image editing	Apache 2.0
Qwen-Image-Layered	Dec 19, 2025	Splits an image into layers	Apache 2.0
Edit-2511	Dec 23, 2025	Better character and group consistency	Apache 2.0
Qwen-Image-2512	Dec 31, 2025	More realistic people, textures and text	Apache 2.0
Qwen-Image-2.0	Feb 10, 2026	Generation and editing in one model, 2K	cloud only
Qwen-Image-3.0	Jul 21, 2026	Flagship: long prompts, small text	cloud only
Qwen-Image-2.1	Sep 20, 2026	7B parameters, transparent backgrounds	non-commercial
Qwen-Image-2.1-Turbo	Oct 9, 2026	2.1 in 8 steps instead of 40	non-commercial

For details on the latest open release, see our news post on Qwen-Image-2.1.

What Qwen Image can do

Long prompts and complex layouts. Qwen-Image-2.0 handled instructions up to 1,000 tokens; Qwen-Image-3.0 handles up to 4,500. In Alibaba’s demo, a 3,700-token prompt produced a 3×3 grid of nine different infographics in a single pass: formulas, diagrams, a comic and bilingual captions. Newspaper pages, storyboards and exam sheets are typical 3.0 jobs.

Small text. Qwen-Image-3.0 renders legible lettering as small as 10 pixels and can reproduce a full page of a research paper with LaTeX formulas: subscripts, fractions and Greek letters.

12 languages and 100+ styles. Version 3.0 renders text in 12 languages, mimics website, game and livestream interfaces and knows over 100 art styles. For fresh data, such as today’s weather forecast, the model can pull from web search.

Generation and editing in one model. Since 2.0 you don’t need a separate Qwen-Image-Edit: the same model swaps backgrounds, adds and removes objects and writes annotations onto photos. Via the API, Qwen-Image-3.0 takes 1–3 input images, while Qwen-Image-2.1 takes up to 10 and edits a region by mask or hand-drawn mark.

Transparent backgrounds and layers. Qwen-Image-2.1 outputs PNG with an alpha channel out of the box. The separate Qwen-Image-Layered splits a finished image into several RGBA layers — 3 or 8, for example — and each layer can be moved, resized and recolored on its own.

Restoration. In Alibaba’s examples, Qwen-Image-3.0 rebuilds missing parts of an old painting in the original style and removes mold stains.

What Qwen Image is good for

• Infographics, slides and cheat sheets — long structured text in a single image.
• Posters, covers and banners with headlines and subheads.
• Storyboards and comics — many panels with speech bubbles in one generation.
• UI mockups — app and website screens for presentations.
• Products on transparent backgrounds — Qwen-Image-2.1 outputs PNG with no cutout needed.
• Self-hosting — the open versions can be run and fine-tuned on your own hardware.

Qwen Image vs. alternatives

	Qwen-Image-3.0	GPT Image 2.5	Nano Banana 2.1	Seedream 5.0 Pro
Arena rank, text-to-image (Oct 24, 2026)	13	1 (Sunburst)	6	12
Resolution	up to 2048×2048	up to 3840 px	up to 4K	up to 2K
Reference images	1–3 (2.1: up to 10)	several	up to 14	up to 10
Prompt length	up to 4,500 tokens	—	—	—
Text	12 languages, from 10 px	multilingual	multilingual	10+ languages
Open weights	no (2.1: yes)	no	no	no

On overall quality Qwen-Image-3.0 sits next to Seedream 5.0 Pro and trails GPT Image 2.5 and Nano Banana 2.1. Its edge is volume: none of its neighbors officially supports a 4,500-token prompt. Among open-weight models, Qwen-Image-2.1 ranks 19th — higher than any open competitor.

How to write prompts for Qwen Image

• Describe the layout, not just the scene. “A 3×2 grid, each panel with an icon and a two-line caption, headline on top” works better than “a nice infographic.”
• Quote text word for word. Write every label exactly as it should appear and add “no other text.”
• Split long prompts into blocks. For 3.0, start with the overall style, then describe each panel in order.
• Write prompts in English or Chinese. Alibaba’s API officially supports these two languages; others get translated with losses.
• Negative prompts are 3.0-only. Qwen-Image-3.0 has a separate field for what must not appear in the image.
• For edits, say what changes and what stays. “Replace the background with a white studio; keep the product, lighting and shadows.”

Can you run Qwen Image locally

Yes, but only the open versions. Weights for Qwen-Image, Edit, 2512, Layered and 2.1 are on Hugging Face and ModelScope, and the models run in Diffusers and ComfyUI. The original 20B Qwen-Image is heavy: with layer-by-layer offloading to system RAM, DiffSynth-Studio runs it on a 4 GB GPU, but generation is slow. Qwen-Image-2.1 is lighter — 7B parameters in the generator plus an 8B Qwen3-VL encoder.

Check the license before you run it. The 2025 releases are Apache 2.0 and fine for commercial use. Qwen-Image-2.1 and Turbo ship under the Qwen Research License, which allows non-commercial use only. Qwen-Image-2.0 and 3.0 can’t be downloaded at all: they run only in Alibaba’s cloud.

How to use Qwen Image in AI2Face

Qwen Image is available in the AI2Face generator — no GPU and no Alibaba Cloud account needed. You can run one prompt on Qwen Image, GPT Image, Nano Banana and Seedream and keep the best result. The cost of each generation is shown before you run it, you can pay by card or crypto, and credits from packs don’t expire. New accounts get welcome credits after signing up. For an overview of other models, see our guide to AI image generators.

Limitations

• The best version is closed. Qwen-Image-3.0 and 2.0 aren’t released, with no weights and no technical report.
• The open 2.1 isn’t for commercial use. The Qwen Research License bans commercial use without a separate agreement with Alibaba.
• Less common scripts aren’t guaranteed. The series is strongest in Chinese and English, and Alibaba doesn’t list 3.0’s 12 languages.
• Resolution tops out at 2K. Large-format print needs upscaling or a 4K model.
• 3.0 takes fewer references. The flagship accepts up to 3 input images, while 2.1 takes up to 10.

Bottom line

Qwen Image is the pick when an image is heavy on text and layout: infographics, slides, storyboards, newspaper pages. Use Qwen-Image-3.0 for those. If you need self-hosting for a non-commercial project, go with the open Qwen-Image-2.1; for commercial self-hosting, Qwen-Image-2512 under Apache 2.0. For photorealism and text in less common scripts, GPT Image 2.5 and Nano Banana 2.1 are stronger — compare them with Qwen Image on the same prompt in AI2Face.