
<p>Qwen Image is Alibaba's family of image foundation models from the Qwen team, built around one idea: a single model that both generates and edits, with text rendering good enough to treat an image as a document. Civitai hosts the API line so you can run it on-site without a Model Studio key.</p>

<p><strong>Built by the <a target="_blank" rel="ugc" href="https://qwen.ai/">Qwen team at Alibaba</a>.</strong> All credit for the models goes to them. Their release write-ups are the canonical source for what each version can do - <a target="_blank" rel="ugc" href="https://qwen.ai/blog?id=qwen-image-3.0">Qwen-Image-3.0</a> and <a target="_blank" rel="ugc" href="https://qwen.ai/blog?id=qwen-image-2.0">Qwen-Image-2.0</a>.</p>

<h3><strong>Versions on Civitai</strong></h3>
<ul>
  <li><p><strong>3.0</strong> - the current release (July 2026). Text-to-image and image editing.</p></li>
  <li><p><strong>2.0</strong> - the 7B unified generation-and-editing model (February 2026). Ranked #1 on the AI Arena ELO leaderboard for both generation and editing at launch.</p></li>
</ul>
<p>Open weights for the earlier Qwen-Image releases live on the <a target="_blank" rel="ugc" href="https://civarchive.com/models/1864281/qwen-image">Qwen-Image</a> page if you want to run them locally.</p>

<h3><strong>Long, detailed prompts</strong></h3>
<p>3.0 accepts instructions up to roughly 4,500 tokens, up from about 1,000 in 2.0. That is the headline change, and it is aimed at composing an entire dense image in one pass rather than generating and then patching: magazine-style layouts, storyboards with recurring characters, teaching diagrams that mix figures and equations, product interface mockups.</p>

<h3><strong>Text rendering</strong></h3>
<p>This is what the line is known for. 3.0 renders text down to around 10 pixels while keeping it legible, and handles 12 languages natively. 2.0 established the pattern with 1k-token typography instructions and native 2K output, and holds up well for posters, comics and infographics.</p>

<h3><strong>Editing</strong></h3>
<p>Both versions are unified models, so editing is the same model as generation rather than a separate pipeline. In practice the text-rendering strength carries straight over: annotating a page in handwriting that matches the paper, restoring faded artwork in the original brushwork, compositing subjects from two reference images into one scene.</p>

<h3><strong>A note on weights</strong></h3>
<p>Qwen-Image 1.0 and 2.0 shipped with open weights. 3.0 did not - it launched through Alibaba Cloud Model Studio and Qwen Chat with no weights, no parameter count, and no technical report. If you are looking for something to download and fine-tune, 3.0 is not it, and the versions on this page run against Alibaba's hosted API rather than on our own workers.</p>

<h3><strong>Links</strong></h3>
<ul>
  <li><p><a target="_blank" rel="ugc" href="https://qwen.ai/blog?id=qwen-image-3.0">Qwen-Image-3.0 release notes</a></p></li>
  <li><p><a target="_blank" rel="ugc" href="https://qwen.ai/blog?id=qwen-image-2.0">Qwen-Image-2.0 release notes</a></p></li>
  <li><p><a target="_blank" rel="ugc" href="https://github.com/QwenLM/Qwen-Image">QwenLM/Qwen-Image on GitHub</a></p></li>
  <li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Qwen">Qwen on Hugging Face</a></p></li>
</ul>


Files with SHA256: f182eaacd3fafa7313fb8a5507297e82b985cf384fb2e6213a90643f33a12515