Qwen Image is a family of models, not a single checkpoint: choose by task, access route, and license. For a combined generation-and-editing workflow, Qwen-Image-2.1 and Qwen-Image-2.1-Turbo share a 7B visual-generation architecture, but Turbo uses a different saved sampling schedule. Qwen lists downloadable weights as well as hosted 2.1 Pro and Turbo APIs.
What each Qwen Image model is for
The original Qwen-Image is a 20B MMDiT text-to-image model; its project describes complex text rendering and image editing among its strengths (Qwen-Image project). For a dedicated editing checkpoint, Qwen-Image-Edit-2511 is a 20B image-to-image model whose card describes reduced image drift, improved character consistency, and support for industrial-design generation and geometric reasoning (model card).
Qwen-Image-Layered is for a different job: it separates an input image into multiple RGBA layers that can be manipulated independently (model card). If you want one checkpoint for both text-to-image generation and editing, Qwen-Image-2.1 combines those tasks. Its repository also lists native RGBA generation and editing, local edits using circles, annotations, or masks, and support for up to 10 reference images (Qwen-Image-2.1 repository).
Qwen-Image-2.1 and Turbo compared
Qwen describes both checkpoints as having a 7B visual-generation component. The documented sampling settings differ: Qwen-Image-2.1 defaults to 40 inference steps, while Turbo has a recommended eight-step schedule saved in the checkpoint. Qwen lists downloadable weights for both and hosted APIs for 2.1 Pro and Turbo.
| Model | Documented task | Visual-generation component | Sampling configuration | Access listed by Qwen |
| Qwen-Image-2.1 | Text-to-image generation and image editing | 7B | 40 steps by default | Downloadable weights; Pro API |
| Qwen-Image-2.1-Turbo | Accelerated text-to-image generation and image editing | 7B; same architecture as 2.1 | Recommended eight-step schedule saved in the checkpoint | Downloadable weights; Turbo API |
Qwen’s 2.1 repository lists the weights and says the Pro and Turbo APIs were live on October 9, 2026. Hosted access and downloadable checkpoints are both listed for the 2.1 family.
The step count is a configuration detail, not a promise of a particular generation time. Turbo’s model card says its saved schedule loads automatically; setting num_inference_steps by itself does not override that schedule. The card specifies Diffusers support for pipeline-configured sampling sigmas, while the 2.1 repository directs users to the latest Diffusers source. If a Turbo workflow behaves unexpectedly, check that dependency and the checkpoint’s saved schedule rather than assuming the parameter alone changed it.
What one local multi-reference test showed
Qwen documents support for up to 10 reference images, but a model’s stated capacity and a particular local workflow’s outcome are different measures. In Comarca IA’s Spanish-language ComfyUI test on Windows 11 with an RTX 3070 and 8 GB of VRAM, editing with five references completed; the seven-reference attempt displayed a config.py error; and the 10-reference run had not finished after more than 10 minutes. The test did not establish the cause of the error.
What the Qwen-Image-2.1 license means
The Qwen-Image-2.1 Research License Agreement limits the grant to non-commercial research or evaluation. It requires a separate commercial license for commercial use of the Materials (license text). For redistribution, the agreement requires including the license, identifying modified files, and retaining its specified attribution notice. It also requires “Built with Qwen” or “Improved using Qwen” in product documentation when the Materials or outputs are used to create, train, fine-tune, or improve a distributed or available AI model.
License terms vary across the family: the Qwen-Image project, Qwen-Image-Edit-2511, and Qwen-Image-Layered pages state Apache 2.0 for those specific models (Qwen-Image, Edit-2511, and Layered). Qwen-Image-2.1-Turbo’s model card names the Qwen Research License Agreement. For a commercial 2.1 deployment, the separate-license requirement is a key part of choosing the model and access route.
