GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
All Posts
Is Qwen Image 3 Open Source? Weights, Local Use and API
2026/10/07

Is Qwen Image 3 Open Source? Weights, Local Use and API

Is Qwen Image 3 open source? No: Alibaba released no 3.0 weights. What you can download instead (Qwen-Image-2.1, research licence) and how to call 3.0 by API.

Is Qwen Image 3 open source? No. As of October 7, 2026, Alibaba has published no downloadable weights for Qwen-Image-3.0 or Qwen-Image-3.0-Pro. The official announcement of July 21, 2026 links no Hugging Face, GitHub or ModelScope repository[1], and the Qwen organization on Hugging Face hosts no 3.0 model[2]. The only way to run it is through an API.

The confusion is understandable. Google's AI Overview answers "No, Qwen Image 3.0 is not open source" for one query, then tells people searching for API pricing that "Qwen Image is open source under the Apache 2.0 license" and can be self-hosted. Both statements refer to different models in the same family. This guide separates them: which Qwen image models have weights, what licence each one carries, what you can run locally today, and how to call Qwen Image 3.0 if local is not an option.

TL;DR

  • Qwen Image 3.0 has no open weights. No 3.0 repository exists in Qwen's Hugging Face or GitHub organizations as of October 7, 2026[2][3].
  • The Apache 2.0 models are the first generation: Qwen-Image, its Edit releases, Qwen-Image-2512 and Qwen-Image-Layered[2].
  • Qwen-Image-2.1 is the newest model you can download (September 20, 2026), but it ships under the Qwen Research License, which allows non-commercial use only[4][5].
  • ComfyUI and GGUF builds exist for 2.1, not for 3.0[6][7].
  • To use 3.0, call qwen-image-3.0 or qwen-image-3.0-pro through Alibaba Cloud Model Studio[8] or through reAPI's Qwen Image 3.0 API.

Is Qwen Image 3 open source? What Alibaba actually published

Alibaba's Qwen team announced Qwen-Image-3.0 on July 21, 2026 as "the third-generation foundational image generation model in the Qwen-Image series"[1]. The post describes prompts of up to 4.5k tokens, text legible at 10px and native rendering of 12 languages. It does not mention weights, a licence or a download, and the page carries no link to Hugging Face, GitHub or ModelScope[1].

Compare that with the Qwen-Image-2.1 post from September 20, 2026. Its header links GitHub, Hugging Face and ModelScope, and its first sentence reads "We are excited to open-source Qwen-Image-2.1"[4]. When Qwen releases weights, it says so and links them.

The repositories confirm the difference. On October 7, 2026, the Hugging Face API listing for the Qwen organization returned ten Qwen-Image repositories, and none of them is a 3.0 model. A Hugging Face search for qwen-image-3.0 across all users returned zero models[2]. The QwenLM GitHub organization has image repositories for Qwen-Image, Qwen-Image-Layered, Qwen-Image-Bench and Qwen-Image-2.1, but none for 3.0[3].

So the answer is a plain no, with a date attached. If Alibaba publishes 3.0 weights later, the place to watch is the Qwen organization on Hugging Face.

Why search engines say Apache 2.0

The Apache 2.0 claim is true for a different model. The first-generation Qwen-Image (on Hugging Face since August 2025) and its follow-ups, Qwen-Image-Edit, Qwen-Image-Edit-2509, Qwen-Image-Edit-2511, Qwen-Image-2512 and Qwen-Image-Layered, are all tagged apache-2.0 on Hugging Face[2]. An answer engine that merges "Qwen Image" into one product will copy that licence onto 3.0.

The same AI Overview also says Qwen-Image 2.0 was released with open weights. We found no Qwen-Image-2.0 repository in the Qwen Hugging Face organization[2]. Alibaba Cloud lists qwen-image-2.0 and qwen-image-2.0-pro as hosted API models[8].

Which Qwen image models you can download

This is the state of the Qwen organization on Hugging Face on October 7, 2026[2]:

ModelDownloadable weightsLicence on the model card
Qwen-Image (1.0)YesApache 2.0
Qwen-Image-Edit, -Edit-2509, -Edit-2511YesApache 2.0
Qwen-Image-2512YesApache 2.0
Qwen-Image-LayeredYesApache 2.0
Qwen-Image-2.0 / 2.0-ProNo (API only)—
Qwen-Image-2.1YesQwen Research License (non-commercial)
Qwen-Image-3.0 / 3.0-ProNo (API only)—

The table answers the question most people actually have, which is whether they can run a recent Qwen image model on their own hardware. They can, as long as it is 2.1 or older and the licence fits their use. For a full comparison of what changed between 2, 2.1 and 3.0, see Qwen Image 2 vs 3.

Running Qwen-Image-2.1 locally

Qwen-Image-2.1 is the closest thing to a local Qwen Image 3. It is a unified text-to-image and editing model with a 7B-parameter visual generation component (32 single-stream DiT layers), up to 10 reference images and native transparent (RGBA) output[4][5]. The official repository runs through a diffusers pipeline (QwenImage21Pipeline)[5].

Check the licence first

This is the part most "open source" summaries miss. The Qwen-Image-2.1 weights are released under the Qwen Research License Agreement, not Apache 2.0. The licence grants rights "FOR NON-COMMERCIAL PURPOSES ONLY" and defines non-commercial as "for research or evaluation purposes only". Commercial use requires a separate licence requested from Alibaba[9]. If you plan to sell images or ship the model inside a product, the downloadable 2.1 weights do not cover you by default.

How big the download is

The official BF16 files on Hugging Face add up to about 33.1 GB: 14.2 GB for the diffusion transformer, 17.5 GB for the text encoder and 1.35 GB for the VAE[5]. Those are file sizes, not a VRAM requirement; the model card shows CPU offload as a memory option[5].

Community packaging makes it smaller and easier to run:

  • ComfyUI. Comfy-Org publishes repackaged Qwen-Image-2.1 files for ComfyUI with text-to-image and image-edit workflow templates[6].
  • GGUF. Unsloth's quantized builds range from a 2.5 GB Q2_K file to a 14.2 GB F16 file, with Q4_K_M at 4.2 GB. Unsloth notes that a GGUF is "the denoiser only", so you still need the VAE and the Qwen3-VL text encoder alongside it[7].

Both carry the same research licence as the original weights[6][7].

What you give up

Qwen-Image-2.1 is a different, smaller model, and the gap shows in preference voting. On Artificial Analysis's text-to-image arena on October 7, 2026, Qwen-Image-2.1 scored an Elo of 1035, against 1075 for Qwen-Image-3.0 and 1088 for Qwen-Image-3.0-Pro[10]. That is crowd preference on general prompts, not a test of 3.0's headline features such as 4.5k-token layout prompts. If those features are why you want Qwen Image 3, 2.1 is a substitute only in name.

Using Qwen Image 3.0 through an API

Alibaba Cloud Model Studio serves two model ids, qwen-image-3.0-pro and qwen-image-3.0, both supporting text-to-image and editing, up to six outputs per call and a maximum resolution of 2048×2048. Alibaba describes qwen-image-3.0 as "a faster version of qwen-image-3.0-pro"[8].

On reAPI the model ids are the same, and the request is a single POST with a Bearer key:

curl https://reapi.ai/api/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-image-3.0",
    "prompt": "A one-page school exam paper titled \"Unit 3 Geometry\", ten numbered questions with diagrams, answer boxes on the right",
    "size": "3:4",
    "resolution": "2K"
  }'

The response returns a task id. Poll GET /api/v1/tasks/{id} until status is completed, then read output.image_urls. Add up to three public image URLs in image_urls and the same endpoint edits instead of generating[11]. You get no model files, but no GPU, no licence negotiation and no 33 GB download either. Billing dimensions for Standard and Pro are covered in Qwen Image 3.0 API pricing.

FAQ

Is the Qwen image free to use?

It depends on the model. The first-generation Qwen-Image family is free to download under Apache 2.0[2]. Qwen-Image-2.1 is free to download but licensed for non-commercial research and evaluation only[9]. Qwen Image 3.0 is a paid API; Alibaba Cloud gives a 10-image free quota per 3.0 model in its Singapore region[12].

Can you run Qwen image locally?

Yes for Qwen-Image 1.0-family models and Qwen-Image-2.1, which have published weights. No for Qwen Image 3.0, which has none as of October 7, 2026[2].

Which Qwen models are open-source?

Among image models: Qwen-Image, Qwen-Image-Edit (including 2509 and 2511), Qwen-Image-2512 and Qwen-Image-Layered under Apache 2.0, plus Qwen-Image-2.1 under the non-commercial Qwen Research License[2]. Qwen-Image-2.0 and 3.0 are API-only.

Is Qwen Image 2.0 open source?

We found no Qwen-Image-2.0 weights in the Qwen Hugging Face organization[2]; Alibaba Cloud offers it as a hosted API model[8]. The open-weight release after 1.0 is Qwen-Image-2.1.

Qwen image 3.0 ComfyUI

There is no local Qwen Image 3.0 model for ComfyUI because no weights exist. ComfyUI's official repackage covers Qwen-Image-2.1[6]. To use 3.0 in a pipeline, call it through an API.

Qwen image 3 GGUF

None exists. A Hugging Face search on October 7, 2026 found GGUF builds for Qwen-Image-2.1 (for example Unsloth's) but no Qwen Image 3.0 repository of any kind[2][7]. A file labelled "Qwen Image 3 GGUF" is either mislabelled or not from Alibaba.

Qwen image 3 model size

Alibaba has not published a parameter count for Qwen Image 3.0; the announcement does not state one[1]. For comparison, Qwen-Image-2.1 has a 7B-parameter visual generation component[4].

Qwen-Image 3 huggingface

There is no Qwen-Image-3.0 page in the Qwen Hugging Face organization as of October 7, 2026[2]. The newest image generation model there is Qwen/Qwen-Image-2.1.

Choosing between local weights and the Qwen Image 3 API

If you need weights on your own machine, pick by licence: the Apache 2.0 Qwen-Image 1.0 family for commercial work, Qwen-Image-2.1 for research and evaluation, or a commercial licence from Alibaba for 2.1. If reAPI's hosted 2.1 is closer to what you need, it is on the Qwen-Image-2.1 model page.

If you need what 3.0 does well, long layout prompts and small legible text, there is no local route today. Is Qwen Image 3 open source? Not as of October 7, 2026, so the practical path is the Qwen Image 3.0 API, with parameters documented in the API docs.

References

  1. Qwen Team. Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge. Retrieved October 2026 from qwen.ai/blog?id=qwen-image-3.0
  2. Hugging Face. Qwen organization model listing (search "Qwen-Image"). Retrieved October 2026 from huggingface.co/Qwen
  3. GitHub. QwenLM organization repositories. Retrieved October 2026 from github.com/QwenLM
  4. Qwen Team. Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation. Retrieved October 2026 from qwen.ai/blog?id=qwen-image-2.1
  5. Qwen. Qwen-Image-2.1 model card and files. Retrieved October 2026 from huggingface.co/Qwen/Qwen-Image-2.1
  6. Comfy-Org. Qwen-Image 2.1 repackaged for ComfyUI. Retrieved October 2026 from huggingface.co/Comfy-Org/Qwen-Image-2.1
  7. Unsloth. Qwen-Image-2.1-GGUF. Retrieved October 2026 from huggingface.co/unsloth/Qwen-Image-2.1-GGUF
  8. Alibaba Cloud. Model Studio: Image generation and editing. Retrieved October 2026 from alibabacloud.com/help/en/model-studio/image-model
  9. Qwen. Qwen Research License Agreement (Qwen-Image-2.1). Retrieved October 2026 from huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE
  10. Artificial Analysis. Text to Image Leaderboard. Retrieved October 2026 from artificialanalysis.ai/image/leaderboard/text-to-image
  11. reAPI. Qwen Image 3.0 API docs. reapi.ai/docs/qwen-image-3
  12. Alibaba Cloud. Model Studio: Model inference pricing. Retrieved October 2026 from alibabacloud.com/help/en/model-studio/model-pricing
All Posts

More Posts