Qwen-Image-2.1.

Qwen-Image-2.1 is an open-weight image model from Alibaba (Qwen team). Specs, recommended settings, hardware and licence terms, from the developer's own pages.

Official model card

Status: Reference
Task: Text to image, Image editing
Input: Prompt, optional reference images (up to 10)
Output: Images at 2K, optionally transparent
Developer: Alibaba (Qwen team)
Parameters: 7B
Licence: Qwen Research License
Released: September 2026

What the model does.

Paraphrased from the official model card.

Qwen-Image-2.1 is the September 2026 Qwen-Image release from Alibaba's Qwen team. One model does both text-to-image and image editing. Its generation transformer has 7B parameters across 32 single-stream DiT layers, with Qwen3-VL 8B as the text encoder, and it generates natively at 2K.

It can generate transparent RGBA images, edit with up to 10 reference images, and take local edits marked with circles, painted annotations or masks. Unlike Qwen-Image-2512, it's released under the Qwen Research License, which doesn't allow commercial use. An 8-step Turbo checkpoint followed in October 2026 under the same licence.

Reference page.

We don't currently serve Qwen-Image-2.1. The specs, settings and licence terms on this page come from Alibaba (Qwen team)'s own pages.

Live for images today: Z-Image Turbo (text to image) and FLUX.2 [klein] 4B (text to image). See Image generation for what the API does now.

Is Qwen-Image-2.1 licensed for commercial use?

Licence: Qwen Research License.

Every model's licence

No, not under its open licence. The Qwen Research License Agreement allows research or evaluation use only, and commercial use needs a separate licence from the Qwen team. Qwen-Image-2.1-Turbo is under the same licence. Earlier Qwen-Image releases such as Qwen-Image-2512 are Apache-2.0.

Qwen Research License Agreement, release date September 20, 2026. Section 2(a) grants use, reproduction, distribution and modification for non-commercial purposes only, which the agreement defines as research or evaluation. Qwen-Image-2.1, LICENSE
Section 2(b): commercial use needs a separate commercial licence from the Qwen team, requested at model-business@notice.qwencloud.com. Qwen-Image-2.1, LICENSE
Section 3: redistribution must include a copy of the agreement, prominent notices on modified files, and a Notice file with the stated attribution. Qwen-Image-2.1, LICENSE
Section 4(b): if you use the materials or their outputs to create, train, fine-tune or improve an AI model that's distributed or made available, you must display "Built with Qwen" or "Improved using Qwen" in its documentation. Section 4(c): "Qwen" can't be the primary name of a derivative. Qwen-Image-2.1, LICENSE
Governed by Chinese law, with the People's Courts in Hangzhou having exclusive jurisdiction. Qwen-Image-2.1, LICENSE
Qwen-Image-2.1-Turbo is under the same Qwen Research License. Qwen-Image-2.1-Turbo model card

This is our reading of the licence, not legal advice. Read the licence text before you rely on it.

Qwen-Image-2.1 hardware requirements.

Published by Alibaba (Qwen team) for running the weights yourself. These are the developer's figures, not Windpaint's.

Every model's published VRAM

Published VRAM: Not stated
Applies to: No VRAM figure on the card or README. bf16 weights; CPU offload shown for limited memory.

Neither the card nor the README gives a VRAM figure. The examples load QwenImage21Pipeline in bfloat16, and for GPUs with limited memory the README shows pipe.enable_model_cpu_offload(). Qwen-Image-2.1 repo
Requirements in the README: torch>=2.4.0, transformers>=5.17, and diffusers installed from source. Qwen-Image-2.1 repo
Architecture: a 32-layer, 7B single-stream DiT; a Qwen3-VL 8B text encoder that encodes both the instructions and any condition images; and a 64-channel RGBA VAE with 16× spatial compression. Qwen-Image-2.1 repo
The README's news section lists day-0 support in ComfyUI (with compatible weights at Comfy-Org/Qwen-Image-2.1 and text-to-image and editing workflow templates), vLLM-Omni, SGLang and LightX2V. These are community projects, not Qwen releases. Qwen-Image-2.1 repo
The README says ModelScope's DiffSynth-Studio (a community tool) supports LoRA training for Qwen-Image-2.1. Qwen-Image-2.1 repo

Qwen-Image-2.1 native resolution.

What the developer's own code and examples use.

Native 2K. Recommended sizes: 2048 × 2048 (1:1), 2400 × 1792 (4:3), 1792 × 2400 (3:4), 2528 × 1696 (3:2), 1696 × 2528 (2:3), 2752 × 1536 (16:9) and 1536 × 2752 (9:16). Qwen-Image-2.1 model card
Both the original and Turbo checkpoints default to 2048 × 2048. Qwen-Image-2.1 repo

Recommended settings for Qwen-Image-2.1.

Steps, guidance and defaults from the developer's card and code, for running the weights yourself.

Default num_inference_steps is 40 for the original checkpoint. The README's SGLang example runs 40 steps with --guidance-scale 1. Qwen-Image-2.1 repo
Scheduler: flow matching with Euler discrete scheduling and dynamic shifting. Qwen-Image-2.1 repo
Qwen-Image-2.1-Turbo, released on October 9, 2026, uses the same 7B architecture and runs 8 denoising steps at CFG 1. Its schedule is saved in the checkpoint and loads automatically; setting num_inference_steps alone doesn't override it. Qwen-Image-2.1 repo
Prefix KV cache: with causal_condition: true (the default), the text and condition-image prefix is computed once at the first step and reused for every later step. Qwen-Image-2.1 repo
Editing takes up to 10 reference images. The card also lists local edits marked with circles, painted annotations or separate masks. Qwen-Image-2.1 model card

Qwen-Image-2.1 prompt guide.

From Alibaba (Qwen team)'s own prompt guide, prompt tooling and examples.

For transparent images, use the recommended format: This is an RGBA image with transparency. <your description>. The image has alpha channel and the background is transparent. Qwen-Image-2.1 model card
The README recommends the official prompt rewriting models for expanding short prompts: two fine-tuned Qwen3.5-VL 9B checkpoints, Qwen/Qwen-Image-2.1-PE-T2I for text-to-image and Qwen/Qwen-Image-2.1-PE-I2I for editing. The rewriter returns a long English prompt and can choose an aspect ratio. Qwen-Image-2.1 repo
The card's example puts the text to render in double quotes: a neon shop sign that reads "QWEN IMAGE 2.1". Qwen-Image-2.1 model card
For multi-reference edits, the README's example passes a list of images with a prompt that places them in one scene: "These three characters are sitting around a campfire in a forest". Qwen-Image-2.1 repo

Questions.

Can I call Qwen-Image-2.1 on Windpaint?

Not at the moment. For images today, z-image-turbo and flux-2-klein-4b are live.

Can I use Qwen-Image-2.1 commercially?

Not under its open licence. The Qwen Research License allows research or evaluation use only, and commercial use needs a separate licence from the Qwen team. Turbo is under the same licence.

Is Qwen-Image-2.1 open source?

No, it's open-weight. The weights are public on Hugging Face and ModelScope, but under the Qwen Research License Agreement, which limits use to non-commercial research and evaluation.

How much VRAM does Qwen-Image-2.1 need?

The card and README don't give a VRAM figure. They load the weights in bfloat16 and show CPU offload for GPUs with limited memory.

What settings does Qwen-Image-2.1 use?

The README's defaults are 40 inference steps at 2048x2048. Qwen-Image-2.1-Turbo uses its saved 8-step schedule at CFG 1.

What resolution does Qwen-Image-2.1 support?

It generates natively at 2K. The recommended sizes run from 2048x2048 for 1:1 to 2752x1536 for 16:9 and 1536x2752 for 9:16.

How do I make a transparent PNG with Qwen-Image-2.1?

Use the card's prompt format: "This is an RGBA image with transparency.", then your description, then "The image has alpha channel and the background is transparent."

Does Qwen-Image-2.1 have a prompt enhancer?

Yes. The Qwen team publishes two prompt rewriting models, Qwen-Image-2.1-PE-T2I for text-to-image and Qwen-Image-2.1-PE-I2I for editing, and recommends using them to expand short prompts.

Does Qwen-Image-2.1 work in ComfyUI?

The official README says ComfyUI supports it natively from day 0, with compatible weights at Comfy-Org/Qwen-Image-2.1 and example workflows for text-to-image and editing.

Qwen-Image-2.1.

Start yours.