Alibaba Shrinks Image AI to One-Third Size

🔥 What happened Alibaba's Qwen team dropped Qwen-Image-2.1, a 7B DiT model that merges text-to-image generation and image editing into one checkpoint. The previous version was 20B and needed a separate edit model. 💡 Why it matters The trick is prefix KV cache reuse: text and reference images are computed once and reused across all denoising steps. That slashes compute, especially with up to 10 reference images. On Qwen's own benchmark it scores 60.28, beating every listed open-weight model – including the 32B FLUX 2 Max at 55.33. ⚡ Our take A 7B model beating a 32B one is a warning shot for anyone still betting on scale. But caveat: it's Qwen's own benchmark – independent tests are still missing.