Qwen open-sources Qwen-Image-2.1, a 7B unified text-to-image and editing model
Qwen released Qwen-Image-2.1, an open-source 7B-parameter text-to-image generation and image editing model. It supports up to 10 reference images, transparent RGBA generation, and 2K resolution. Weights are available on HuggingFace and ModelScope.
China context
- Original name
- 通义千问图像2.1
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate Qwen-Image-2.1 via Diffusers, ComfyUI, or vLLM-Omni with Day-0 support, enabling self-hosted image generation and editing without API costs.
- For investors
- The release of open weights and hosted APIs by Alibaba indicates continued investment in generative AI, potentially pressuring competitors to match capabilities or pricing.
Qwen has open-sourced Qwen-Image-2.1, a unified text-to-image generation and image editing model with a 7B-parameter visual generation component using 32 Single-Stream DiT layers. The model supports native transparent (RGBA) image generation, editing with up to 10 reference images, local edits via circles or masks, and 2K resolution output. It is available on HuggingFace and ModelScope, with Day-0 support from Diffusers, ComfyUI, vLLM-Omni, SGLang, LightX2V, and Intel OpenVINO. A Turbo variant using 8 denoising steps was released on 2026-10-09, along with Pro and Turbo APIs on Alibaba Cloud Model Studio.
Qwen-Image-2.1 uses a lightweight architecture with mixed-granularity attention and prefix KV cache reuse to balance quality and efficiency. It supports up to 10 reference images for multi-subject composition and native RGBA output. The Turbo variant reduces denoising steps to 8 while using the same 7B architecture. Prompt rewriting models based on Qwen3.5-VL 9B are provided for text-to-image and image editing.
Developers outside China can now use a capable open-weights image model with transparent image generation and multi-reference editing, reducing reliance on proprietary APIs. The Day-0 support from major inference frameworks lowers integration cost for self-hosted deployments.
Open weights allow businesses to self-host image generation and editing without per-image API fees, potentially lowering costs for high-volume use cases. The Turbo variant offers faster inference, which may reduce compute costs for real-time applications.
A measurable signal is the number of community fine-tunes and integrations published on HuggingFace and GitHub in the next quarter. Another observable indicator is the adoption rate of the Turbo API on Alibaba Cloud Model Studio, which can be tracked through usage metrics or customer case studies.