Overview
Qwen-Image-2.1 is an open-source image model released by the Qwen team on September 20. A single 7B checkpoint handles both text-to-image and instruction-based editing, with native support for generating and editing RGBA images with alpha channels. It accepts up to 10 reference images per run, and local edits can be scoped with circles, scribbles or separate masks; mixed-granularity attention and KV cache reuse keep the compute cost in check. On September 23 it scored 1,367 on the Arena Image Edit leaderboard, first among open models and 16th overall, just three points behind GPT-Image-1.5-high-fidelity, and it also tops the open-source column of the text-to-image board. vLLM-Omni, SGLang-Diffusion and ComfyUI shipped support on day one, and the Unsloth GGUF build runs locally on a 12GB consumer GPU. The weights ship under a strict non-commercial license.
Key Features
- Unified generation and editing: Text-to-image and instruction editing share one 7B checkpoint, so there is no need to swap models per task.
- Native transparency: Outputs RGBA images with alpha channels directly, so transparent assets need no background removal.
- Three ways to scope edits: Up to 10 reference images, with circles, scribbles or separate masks to limit changes; untouched regions stay as they are.
- Low deployment bar: Mixed-granularity attention and KV cache reuse cut inference cost; the GGUF build runs on a 12GB GPU, with ComfyUI, vLLM-Omni and SGLang-Diffusion already adapted.
Use Cases
- E-commerce and design assets: generate transparent-background product shots and stickers without a matting step
- Reference-based editing: swap backgrounds or elements on product photos while untouched areas stay intact
- Local deployment: ComfyUI workflows and GGUF quantization run on personal GPUs and workstations
Pros
- 1,367 on the Arena Image Edit board, first among open models and 16th overall, three points from the closed-source leader
- The 7B size is deployment-friendly and runs locally on a 12GB consumer GPU
- Native RGBA output removes post-processing for transparent scenes
- Fast ecosystem adoption: ComfyUI, vLLM-Omni and SGLang-Diffusion were available on day one
Pricing
Weights are open on Hugging Face under a strict non-commercial license, so personal and internal use is free to self-host; commercial projects should review the license terms carefully and are better served by official API channels such as Alibaba Cloud Model Studio, billed per call. Refer to official documentation for exact terms and pricing.
Summary
Qwen-Image-2.1 raises the bar for open-source image models: one 7B checkpoint covers both generation and editing, with native transparency and three ways to scope local edits. First place among open models on both Arena image boards, plus a 12GB local deployment floor, makes it one of the most compelling open image models to self-host right now. The main catch is the license: strict non-commercial terms push commercial work to the official API.
Version History
- Qwen-Image-2.1 takes the top open-source spot in both Arena image leaderboards (2026-09-23): Qwen-Image-2.1 scored 1,367 points in the Arena Image Edit leaderboard, first among open models and 16th overall, three points behind GPT-Image-1.5-high-fidelity, and also ranks first among open models for text-to-image. It is a 7B single checkpoint that handles both generation and instruction-based editing, outputs native RGBA transparency and accepts up to 10 reference images. vLLM-Omni, SGLang-Diffusion and ComfyUI all added day-zero support, and Unsloth GGUFs let it run locally on 12GB of VRAM. Weights ship under a strict non-commercial license.
- Qwen open-sources Qwen-Image-2.1: a 7B model unifying generation and editing with native transparency (2026-09-20): The Qwen team open-sourced Qwen-Image-2.1, unifying text-to-image and image editing in a single 7B checkpoint with native transparent-image generation and editing, including RGBA output with an alpha channel. It takes up to 10 reference images per instruction edit, supports circle, scribble and separate-mask local edits, and speeds up inference through mixed-granularity attention and KV cache reuse. Native 2K output is supported across panoramas, infographics and storyboards.