Alibaba launches Qwen-Image-2.1: unified 7B model for image generation and editing

A third of the size, twice the jobs

Alibaba's Qwen team has released Qwen-Image-2.1, an open-weight model that merges text-to-image generation and editing into a single 7-billion-parameter checkpoint for the first time. The previous version, Qwen-Image from August 2025, weighed 20B under an Apache 2.0 license and kept editing in a separate model. Now both tasks fit in one model at roughly a third the volume. The team calls it the most balanced and economical entry in the series, with a catch: the 7B count covers only the diffusion transformer. The full pipeline also loads an 8B Qwen3-VL encoder, so the real memory footprint is larger.

Architecture: a cache that stays put

At the core sits a 32-layer single-stream diffusion transformer with block-causal attention. The practical novelty is in the attention mask: text tokens receive a token-level causal mask, while image tokens get a chunk-level bidirectional mask inside each reference image. Qwen labels this "mixed-granularity attention." The conditional prefix — text plus reference images — sits before the noisy latent and never attends to it; therefore its keys and values stay fixed across every denoising step. The model computes them once, at the first step, and reuses that KV cache for all subsequent steps. The savings scale with the number of reference images, which explains the claimed speed-up in multi-image scenarios.

Capabilities: native transparency, multi-reference, local control

The model outputs native RGBA — images with a built-in alpha channel — from text, edits transparent layers, and extracts objects from photographs; the README recommends a fixed prompt template for transparent output. Multi-reference editing supports up to 10 source images, with examples such as a group photo assembled from 6 portraits or an outfit composed from 5 references. Local control works via circles, drawn annotations, or separate masks while preserving the identity of people and products. Default resolution is 2048×2048 with seven supported aspect ratios up to 2752×1536. On the aesthetic side the team reports better typography, portrait lighting, and fine detail, with emphasis on panoramas, infographics, storyboards, and virtual try-on.

In-house benchmark: leads the open field, still trails closed models

The official comparison runs on Qwen-Image-Bench, the team's internal benchmark. There Qwen-Image-2.1 scores 60.28 overall, above Nano Banana 2.0 at 59.82 and every other open-weight model on the list. FLUX 2 Max, a 32B open model, sits at 55.33. Six closed models rank higher, led by GPT Image 2.5 Sunburst at 67.01. Because the benchmark belongs to the developer, the numbers need independent verification before anyone claims practical superiority.

Deployment: day-zero ecosystem, separate license for production

Immediate (day-zero) support exists in Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V. The minimal Diffusers snippet requires PyTorch 2.4+, Transformers 5.17+, Diffusers from source, Accelerate, and Pillow; the same pipeline handles editing when you pass an image parameter with one or more references. For smaller GPUs, pipe.enable_model_cpu_offload eases memory pressure. For serving, vLLM-Omni adds FP8 quantization, prefix KV caching, CUDA Graph decoding, and tensor parallelism; SGLang adds Cache-DiT, CUDA graphs, multi-GPU, and component offloading. ComfyUI ships with native nodes and converted weights. Commercial deployment requires a separate license from Qwen — the model is not Apache 2.0 like its predecessor.