Qwen-Image-2.1 is free in the way that matters most: the weights are public. Alibaba released the 7B unified generation-and-editing model on September 20, 2026 under the Qwen Research License, with native transparent output and editing from up to 10 reference images. There’s no per-image bill and no API key to buy. What you pay with is either a queue (hosted demos) or a GPU (running it yourself). This guide lists every no-cost route, what each one gives you, and what the license lets you do with the output.
If you need the feature overview first, read What is Qwen-Image-2.1. If you’re past trying and want code, the how-to guide has the diffusers examples and a wrapper you can test in Apidog.
The free routes at a glance
| Route | Cost | Best for | Limits |
|---|---|---|---|
| Qwen Chat, “Image Generation” | Free account | Fastest first look, prompt rewriting built in | Consumer caps, no seed control, model behind the UI can change [VERIFY that 2.1 is the model serving Qwen Chat] |
| Hugging Face Space | Free (ZeroGPU queue) | Testing the exact 2.1 checkpoint with references and transparency | Queue times, per-user GPU quota, no batch |
| ModelScope Studio | Free | Same as the Space, better throughput from mainland China | Account needed |
| ComfyUI on your GPU | Free software, your hardware | Repeatable workflows, transparent layers, local editing with masks | GPU memory; setup |
| diffusers on your GPU or a cloud notebook | Free software, your hardware or a free-tier notebook | Programmatic use, seeds, automation | Same; free notebooks may not fit bf16 |
Route 1: Qwen Chat
The quickest way to see what 2.1 does is the Image Generation mode in Qwen Chat, which Qwen points to from its own landing page. Type a short request in any language; the chat layer expands it into a long structured prompt (the same job the published PE-T2I prompt-rewriting model does) and returns an image. Editing works by attaching an image and describing the change.
Good for: judging typography, portrait quality, and whether the editing style suits you. Not good for: reproducibility. There’s no seed, the caps aren’t published, and consumer chat products swap models without notice, so confirm which model is answering before you draw conclusions about 2.1 specifically.
Route 2: the Hugging Face Space
The official Qwen-Image-2.1 Space runs the actual checkpoint on shared GPUs. It exposes the features the launch post shows: text-to-image, editing with reference images, and transparent output when you ask for it. Log in to a free Hugging Face account for a larger ZeroGPU quota; anonymous users get less and wait longer.
Three tests worth running here before you spend any GPU money of your own:
- Transparency. Prompt: “This is an RGBA image with transparency. A cartoon fox sticker. The image has alpha channel and the background is transparent.” Download the PNG and open it over a checkerboard. Zoom into fur edges; that’s where background removers fail and native alpha should not.
- Multiple references. Upload three photos of one product from different angles and ask for a lifestyle shot. Check that the label text and shape survive.
- Text rendering. A sign or poster with a quoted string in two scripts. Qwen’s edge is here, and it’s cheap to confirm.
Every Gradio Space also exposes its functions as HTTP endpoints, which means you can script these tests. The gradio_client package or a direct call to the Space’s /gradio_api/call/<fn> route lets you send a prompt and fetch the result file; document the exact route from the Space’s “Use via API” link [VERIFY route name for this Space]. Once it’s an HTTP call, you can build it as a request in Apidog, assert on the returned file’s type and size, and rerun the same three tests against your own server later without rewriting anything.
Route 3: ModelScope
Qwen mirrors the weights and the demo on ModelScope. If Hugging Face is slow or throttled where you are, the Studio demo there runs the same model, and the model page is the fastest download mirror for users in mainland China.
Route 4: ComfyUI on your own GPU
ComfyUI added native support on launch day with a template workflow. The setup is four steps: update ComfyUI, download the weights from Hugging Face, load the template, and run it with prompts and reference images. Transparent output and multi-reference editing are wired into the template, and you can save the whole graph as a JSON file, which makes ComfyUI the most repeatable free route short of writing code.
Qwen hasn’t published VRAM requirements. The reference stack is a 7B generator, an 8B Qwen3-VL text encoder, and a VAE, in bfloat16, at 2048 x 2048 and 40 steps. Budget for a high-end consumer or data-center GPU for the plain bf16 path. Lighter paths exist: enable_model_cpu_offload() in diffusers, FP8 through vLLM-Omni, and LightX2V acceleration, all listed in the GitHub README. Community FP8 checkpoints for ComfyUI usually appear within days of a Qwen release; check the model’s Hugging Face “Spaces and models using this” list before converting your own.
Route 5: diffusers, locally or in a free notebook
The reference code is short. Install torch>=2.4.0, transformers>=5.17, accelerate, pillow, and diffusers from GitHub main, then:
import torch
from diffusers import QwenImage21Pipeline
pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload() # skip and .to("cuda") if the whole model fits
image = pipe(
prompt="This is an RGBA image with transparency. A paper plane icon. The image has alpha channel and the background is transparent.",
num_inference_steps=40,
generator=torch.Generator("cuda").manual_seed(7),
).images[0]
assert image.mode == "RGBA"
image.save("plane.png")
Free cloud notebooks give you a GPU for a few hours at a time; whether the bf16 pipeline fits their memory depends on the card you’re assigned, which is why the offload line is there. The how-to guide extends this into editing with references and a small HTTP server.
What “free” allows: the license
Free to download is not free to sell. Qwen-Image-2.1 ships under the Qwen Research License Agreement. The clause that matters: you “shall not use the Materials for any commercial purpose without obtaining a separate commercial license” from Qwen. Redistributed weights or fine-tunes must keep the license, add a “Built with Qwen” notice, and can’t use “Qwen” as their primary name.
So every route above is fine for learning, research, portfolio work, internal experiments, and deciding whether to license it. Selling the images, or putting the model behind a paid product, needs the commercial agreement first. That is the biggest difference from the Apache-licensed original Qwen-Image, and from paid APIs like Nano Banana 2 or gpt-image-2.5, whose per-image price includes commercial rights. If you need commercial output without a license conversation, Alibaba’s hosted Qwen Image 3.0 at $0.03 per image is the intended path; the 2.1 vs 3.0 comparison lays out when each makes sense.
Free but not fast: what to expect
- Hosted demos queue. At launch week, expect minutes per image on the Space at peak US hours. The quota resets daily for logged-in users.
- 2K at 40 steps is slow on consumer GPUs. Test at 1024-class sizes first, then scale up once the prompt is right.
- Multi-reference edits are the case 2.1 optimized: reference images are encoded once and cached across steps, so five references don’t cost five times one.
- The chat route hides the seed. If you need the same image twice, use the Space, ComfyUI, or diffusers, where you control it.
Turn your free experiments into a test suite
Once one route works, the risk is that your prompts drift while the model stays the same, or the model changes while your prompts stay the same, and you can’t tell which. Whether you’re calling the Space’s Gradio endpoint or your own wrapper, define the request once in Apidog: the prompt, the seed, the transparency flag, and the reference files. Save the transparency test and the three-reference edit as test cases with assertions on the response type and size, and rerun them whenever you change anything. When you graduate from free to a licensed deployment or to the 3.0 API, the suite moves with you; only the base URL changes. Download Apidog to set that up.
FAQ
Is Qwen-Image-2.1 completely free? The weights and the demos are free. Commercial use is not included in the license; you’d need a separate agreement from Qwen.
Which free route runs the real 2.1 model? The Hugging Face Space, ModelScope Studio, ComfyUI, and diffusers all run the released checkpoint. Qwen Chat likely does too, but consumer chat products don’t guarantee which model version answers.
Can I get transparent PNGs for free? Yes. Ask for an RGBA image in the prompt on the Space, in ComfyUI, or in diffusers, and check the file’s mode.
Do I need a GPU? Not for the Space, ModelScope, or Qwen Chat. For ComfyUI or diffusers you need a GPU; Qwen doesn’t publish minimum VRAM, and CPU offload plus FP8 builds lower the bar.
Is there a free API for Qwen-Image-2.1? Not a hosted one from Alibaba. The Space’s Gradio endpoint works for scripting small tests, and the how-to guide shows how to run your own.
Where to go next
Start on the Space with the three tests above, move to ComfyUI or diffusers when you need seeds and repeatability, and read the license before anything ships. The how-to guide is the next step for code, the comparison with 3.0 for the buy-vs-run decision, and Apidog keeps your prompts and assertions stable across all of them.
