Alibaba open-sources Qwen-Image-2.1: 7B image model tops open-source at 60.28
机器之心 · wechat · 2026-09-20
Alibaba's Qwen team open-sourced Qwen-Image-2.1, unifying text-to-image generation and editing in one pipeline aimed at real design workflows.
- Benchmarks: 60.28 on Qwen-Image-Bench, beating NanoBanana2.0 (59.82) and GPT Image 1.5 (59.65) as the top open-source model, with only 7B vision-generation parameters (20-layer Single-Stream DiT, native 2K).
- Capabilities: native RGBA transparent-image generation and subject extraction from photos; up to 10 reference images for multi-image editing with lasso/brush/mask-based local edits; stronger identity and product-packaging fidelity; better text rendering; panorama, infographic, three-view and storyboard generation.
- Efficiency: hybrid-granularity attention — token-level causal masks for text, chunk-level masks for images with KVCache reusing static context — cutting memory and boosting inference on multi-image inputs.
- Weights are on Hugging Face and ModelScope; note the strict non-commercial qwen-research license with no revenue cap.
More from Multimodal
- Indie Britpop AI music video built with Nano Banana Pro, Kling, and free lip-sync via WAN2GP + LTX 2.5 — cocktailpeanut · 2026-11-12
- P2Echo: one text-prompted model segments multi-view echocardiography with high accuracy and low cost — UBC-V · 2026-11-12
- Redditor makes fast-motion short film with H3 and ships Longshot Studio v0.18.5 with face swap — R34vspec · 2026-11-12
- Reimagining Knight Rider's opening as a Netflix drama with Claude Fable 5.2 — justin_hart · 2026-11-12
- Gemini 4 Carbon builds a full 3D aquarium in Blender and Three.js in ~6 hours — sunjiao123sun_ · 2026-11-12
- Awaken Open-Sourced: Turn a Half-Duplex Speech LLM Full-Duplex With Just a Small LoRA — alexcovo_eth · 2026-11-12