Qwen Image 3.0: What It Is and Where You Can Run It

Qwen Image 3.0 is the third generation of Alibaba's Qwen-Image family, and the headline numbers in its launch announcement are a 4.5k token prompt window, legible text rendered as small as 10 pixels, and native rendering across 12 languages. The bigger fact is missing from those numbers: unlike every Qwen image model that shipped in 2025, Qwen-Image-3.0 arrived with no downloadable weights, no license text, no parameter count, and no technical report. If you came here to run it locally, that is the answer, and the rest of this guide covers what you can run instead.

One disclosure up front: Dream Pixel Forge does not ship any Qwen image model. We run Nano Banana, GPT Image 2, and Grok Imagine, plus FLUX.1 Kontext Pro for edits. Nothing here comes from our own testing of Qwen-Image-3.0, it comes from Qwen's announcement, Alibaba Cloud's Model Studio documentation, the QwenLM repositories, and Hugging Face model cards.

What Qwen Image 3.0 is

Qwen frames the release around a single word, "Real" (实), positioned as a step past the earlier generations: Qwen-Image-1.0 was pitched on "Precision" and Qwen-Image-2.0 on "Precision, Variety, Completeness, Beauty, and Authenticity." The stated goal is a shift from good-looking to useful, or in the announcement's own framing, making image generation "a truly deployable productivity tool." Three capabilities carry that claim.

Rich content, meaning very long prompts. The model accepts up to 4.5k tokens of instruction. Qwen's showcase example is a 3x3 grid where every cell is a separate complex infographic, a tunnel safety comic, a spatial geometry lesson, physics projectile motion, a parasitology explainer, the Sylow theorems, a bank internal-control diagram, and more, described in a single 3.7k-token prompt and generated in one pass rather than stitched from nine images. A second demo nests four interfaces inside each other from one instruction: a VSCode window containing a Qwen Chat window containing a WeChat window containing a coffee poster.

Authentic details, meaning small text and micro-texture. Qwen claims precise rendering of text down to 10px, and demonstrates it on a whale shark infographic, a full page of an algebraic geometry paper with multi-line LaTeX derivations, and a newspaper layout. The same section shows handwritten red annotations overlaid onto a book page as an editing task, and a damaged ink painting restored with the original brushwork preserved.

Deep knowledge, meaning world knowledge and UI mimicry. The announcement claims coverage of 12 languages, multiple fonts, over 100 artistic styles, and a range of real UI interfaces, with Japanese, Korean, and Spanish examples shown. It also states the model can connect to the internet to retrieve current information, illustrated with a weather forecast image for a specific date.

All of the above are vendor demonstrations. There is no third-party evaluation of Qwen-Image-3.0 to check them against: Qwen published no benchmark scores with the release, and with API access invite-gated nobody has run it at evaluation scale. Read them as claims, not measurements.

What changed against the earlier Qwen image models

Eight releases in twelve months, and the versions are not a straight line. Through 2025 Qwen ran two parallel tracks, generation and editing, then merged them into the unified 2.0 model. This table is assembled from the news log in the QwenLM/Qwen-Image repository and the release blogs.

Measured against Qwen-Image-2.0 specifically, the concrete deltas are the prompt window going from 1k to 4.5k tokens and the small-text and multilingual claims. Everything else in the announcement is a demonstration rather than a spec. Note also that 2.0, not 3.0, was where the open-weight break happened: the repository's news entry for 2.0 points readers to Qwen Chat and carries none of the "we released the weights" language that accompanies every 2025 release. Qwen-Image-3.0 continues that pattern rather than starting it.

Is Qwen Image 3.0 open source, and what is the license?

Qwen-Image-3.0 is not open source and has no published license. There is no Qwen/Qwen-Image-3.0 repository on Hugging Face or ModelScope, no model card, no parameter count, and no arXiv technical report. The QwenLM/Qwen-Image README does not mention 3.0 at all. As of this writing there is also no official statement committing to a future weights release.

The models that are open are all Apache 2.0, which is genuinely permissive, commercial use included, and that license is stated on each model card:

• Qwen/Qwen-Image, 20.4B parameters, Apache 2.0, with a technical report
• Qwen/Qwen-Image-2512, Apache 2.0
• Qwen/Qwen-Image-Edit-2511, Apache 2.0
• Qwen/Qwen-Image-Layered, Apache 2.0

So the practical position is a split lineup. The open branch stopped at the December 2025 models, and the flagship branch, 2.0 and 3.0, is a hosted product. If your reason for choosing Qwen was that you could download it, the model you can download is Qwen-Image-2512 or Qwen-Image-Edit-2511, not 3.0.

Where the Qwen AI image generator runs today

Two places, and one of them is gated.

Qwen Chat. The announcement's only call to action is chat.qwen.ai with the text-to-image feature selected. This is the free surface, and with the API invite-gated it is the only one most people can reach.

Alibaba Cloud Model Studio. The API model id is qwen-image-3.0-pro, and Alibaba's model selection page marks it "currently in invite-only testing" in the International region. Its documented envelope: text-to-image and editing under the same model id, up to 6 images per call, maximum resolution 2048x2048, recommended for complex layout generation, small-text rendering, and multilingual font support. Pricing is listed as "Limited-time free," against $0.075 per image for qwen-image-2.0-pro and $0.035 for qwen-image-2.0.

Worth flagging: Alibaba's Qwen image API reference has not caught up with the launch. It does not list qwen-image-3.0-pro among its models, and the prompt ceiling it documents is 1,300 tokens for the 2.0 series and 800 for the rest. The 4.5k figure appears in the announcement and nowhere in the API docs yet. If you get invite access, verify the real limit against the endpoint rather than the blog post.

Everywhere else carries the older, open Qwen models only. Replicate's Qwen collection, NVIDIA's build catalog, and the ComfyUI ecosystem all host the 2025 family. None of them has 3.0, and none of them can until weights exist.

Running Qwen image models locally on ComfyUI

This section is about the open models, since 3.0 is not among them. ComfyUI supports Qwen-Image and Qwen-Image-Edit natively, with an official workflow tutorial rather than a custom-node scramble.

The constraint is memory. At 20B parameters, the fp8 build (weights stored at 8-bit precision to halve memory) in ComfyUI's own tutorial consumes roughly 86 percent of a 24GB card, so a 24GB GPU is the comfortable floor for fp8. Anything smaller means community GGUF quantizations, the same trick pushed further, at a visible quality cost. Weights come from Hugging Face or ModelScope, and third-party acceleration projects such as Qwen-Image-Lightning publish distilled variants that cut step counts. If your interest in Qwen is specifically local, unlicensed, offline generation, the December 2025 models are a real and capable option, and they are the ones to plan around.

How to write Qwen image prompts

Qwen published no prompting guide with 3.0. What follows is drawn from the documented API surface for the Qwen image family plus what the announcement's own examples demonstrate, so treat it as informed rather than tested.

Negative prompts exist here, unlike most current models. The Qwen image API takes a real negative_prompt parameter, and Alibaba's docs list "to use a negative prompt to exclude specific elements" as a reason to pick Qwen over its sibling families. Alibaba's own example negative is a quality laundry list: low resolution, distorted limbs, malformed fingers, oversaturated colors, wax-figure appearance, excessive smoothness, chaotic composition, blurry or warped text. This is a genuine dialect difference. FLUX.2 has no negative field at all, and on Nano Banana, GPT Image 2, and Grok Imagine we get better results describing the state we want than the state we do not.

There is an LLM rewriting your prompt by default. The prompt_extend parameter defaults to true and silently expands a short prompt into a long, styled one. Alibaba's documented example turns a one-line cat description into a full paragraph specifying golden-hour lighting, a blurred autumn garden, Canon EOS R5, f/1.2, and 8K. That is fine for exploration and hostile to control. If you are matching a brand or reproducing a layout, set it to false and write the detail yourself, and note that only asynchronous calls return the rewritten prompt, so with the default on you often cannot see what the model was actually told.

Spend the long window on layout, not adjectives. Every capability Qwen demonstrates at 3.7k tokens is structural: which panel holds what, where each label sits, what each nested interface contains. A 4.5k window filled with style words buys nothing. Describe regions in reading order, name the artifact type for each one, and give exact strings for text.

Quote the exact text you want rendered. This holds on every text-capable model, Qwen included. Settle the wording before you prompt the visual, because a model asked to invent a slogan renders it worse than one handed a fixed string. Alibaba's own poster example writes each headline, subtitle, and speech bubble out in full inside the prompt.

Two more parameters worth knowing before you build anything on it: output is PNG, and generated image URLs expire 24 hours after the call, so download immediately. There is also a watermark flag, off by default, that stamps a Qwen-Image mark in the bottom-right corner.

Qwen Image 3.0 compared to the models we run

Prompting habits are model-specific, and the useful comparison is which model to reach for, not which is best. Our side of the table is drawn from operating three models daily, the Qwen side from documentation.

On measured quality there is one data point, and it is Qwen's own. Qwen-Image-Bench is an Apache 2.0 evaluation toolkit Qwen publishes, scoring 1,000 prompts across five dimensions and 56 facets with a fine-tuned judge model. Its top-five table reads: GPT Image 2 at 64.69 overall, Nano Banana 2.0 at 59.82, GPT Image 1.5 at 59.65, Nano Banana Pro at 59.45, and Qwen Image 2.0 Pro at 57.84. That is Qwen's own benchmark placing Qwen fifth, which is a point in its favor as a benchmark. It also predates 3.0, so it tells you where the previous flagship sat, not where this one does.

Per-model prompting detail for the three we operate lives in the sibling guides: Nano Banana prompts, GPT Image 2, and Grok Imagine prompts. The open-weight comparison most people are actually running is Qwen against Flux, covered in Flux prompts, and if you need a wider style vocabulary for any of them, AI art styles covers around thirty registers with prompt scaffolding.

If what you want is to compare models rather than commit to one, Dream Pixel Forge runs Nano Banana, GPT Image 2, and Grok Imagine in one studio on a shared credit balance, so you can put the same brief through all three without opening three accounts. Qwen is not among them. For Qwen specifically, Qwen Chat is the free path to 3.0, Hugging Face is where the open 2025 weights live, and Model Studio is where the API sits once the invite gate opens.

The bottom line

Qwen Image 3.0 is a hosted flagship aimed squarely at dense, text-heavy, multilingual production work: newspapers, storyboards, exam papers, infographics, UI mockups. The 4.5k prompt window and the 10px text claim are the real differentiators, and they point at layout jobs rather than pretty pictures. What it is not is the open-weight Qwen people spent 2025 downloading. That branch is still there, still Apache 2.0, and still worth running locally, but it ends at the December 2025 models. Until Qwen publishes weights, a license, or a technical report for 3.0, evaluate it as a product you rent, not a model you own.