Qwen Image 2.1: Open-Source Release Details
Qwen released Qwen Image 2.1 on September 20, 2026. This open-source model features a 7B parameter visual component and supports transparent image generation.
Release Announcement and Core Specifications
The original announcement date for this release was September 20, 2026. On this date, Qwen confirmed the open-sourcing of Qwen Image 2.1, an image model within the Qwen family. The official summary states that the model balances generation quality, inference efficiency, and cost.
The source explicitly identifies the visual generation component as having 7B parameters. The model accepts text input and produces image output. It is described as unifying text-to-image generation and image editing within a single model architecture.
Documented Features and Transparency Support
A specific feature highlighted in the release notes is native support for generating and editing transparent images. The source indicates that this capability is built into the model rather than requiring external processing steps.
The announcement mentions four key improvements introduced in this update. However, the provided source facts do not list the specific details of these four improvements beyond the general description of balancing quality, efficiency, and cost. Therefore, only the transparency support and the unified generation/editing structure are confirmed by the available text.
- Open-source status confirmed.
- Visual generation component size is 7B parameters.
- Supports both text-to-image generation and image editing.
- Native support for transparent image generation and editing.
Evaluation Procedure for Engineers
Engineers can evaluate the model by testing its performance against specific criteria derived from the documented features. First, test the model’s ability to generate transparent images by providing prompts that require alpha channel output. Measure the consistency of the transparency across different prompt variations.
Next, validate the unified nature of the model by comparing the output quality of generated images versus edited images using the same base asset. Compare the inference time for both tasks to assess the claimed balance between efficiency and quality. Since the source does not provide benchmark scores, engineers must measure local or hosted performance independently to determine if the 7B parameter count aligns with their hardware constraints.
Primary sources
Frequently asked questions
Is Qwen Image 2.1 available as open weights?
Yes, the source confirms that Qwen Image 2.1 is open-source. The model card is listed on Hugging Face, indicating public availability of the weights.
What is the parameter count of the visual component?
The visual generation component contains 7B parameters, as stated in the official release summary.