阿里开源 70 亿参数图像模型 Qwen-Image-2.1，官方称对标闭源巨头

Ali 7 Billion Parameter Image Model Qwen-Image-2.1 Open Source Release, Claiming to Challenge Closed-Source Giants with a Slimmer Body AIbase基地 Published inAI News · 4 min read · Sep 21, 2026 6 The Alibaba Qwen team has recently unveiled a remarkable offering to the open-source community: the Qwen-Image-2.1, an image generation and editing model with open weights, has officially launched. What surprised many was its compact size—the visual generation component uses only 7 billion parameters but outperformed most closed-source models on its own evaluation benchmark. The team is confident in their claim, though independent third-party evaluations are still in progress, and the true value of this achievement will need time to prove. More impressively, it doesn't require high-end equipment—just a mainstream consumer-grade GPU like the RTX 3090 can run it smoothly. Qwen-Image-2.1's strongest feature is its ability to "group similar elements." In official demonstrations, each face in a group photo comes from separate individual images. The model acts like a patient puzzle solver, seamlessly merging them into a single composite image. This is made possible by its native support for transparent images—generated images are not standard RGB images but RGBA format with an alpha channel, allowing users to extract objects separately and directly edit text on transparent layers, saving time on layout and post-processing. The model's ability to handle reference images is equally impressive. It can process up to ten reference images at once, supporting tasks such as group image synthesis, virtual try-on, and room design within the same framework. If you want to make local modifications, there's no need to redraw the entire image. Simply draw a circle, apply a mask, or mark a few points, and the model knows exactly which part to modify. The team revealed that architectural changes and the reuse of KV cache have significantly improved inference speed, especially when dealing with a larger number of reference images, where the speed improvement is even more noticeable. The barrier to entry is also very low. The model is now available on Hugging Face, GitHub, and the ModelScope (Magic Mall) community, along with an online demo for users to experiment with. The only thing to note is the licensing: it is released under a research license, which explicitly prohibits commercial use. Enterprises wishing to implement it in production must apply separately for a commercial license from the Qwen team. In other words, personal experimentation and academic exploration are fully open, but if you plan to make money, you must first pass the licensing requirement. This article is from AIbase Daily Scan to view Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications. —— Created by the AIbase Daily Team © Copyright AIbase Base 2024, Click to View Source -