Qwen-Image-2.1: 7B open weights for generation, editing and alpha layers, under a research-only license (2026)

Alibaba's Qwen team released Qwen-Image-2.1 on September 20: one image model that both generates and edits, and the first in the Qwen-Image line to handle transparent backgrounds natively. 1 The weights are on Hugging Face and ModelScope, with day-zero support in diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V. 2 3

The terms are where the release narrows. Qwen calls it open weights, but the weights ship under the Qwen Research License Agreement, which grants use "for research or evaluation purposes only" and says a commercial use needs a separate license. 4 On Qwen's own Qwen-Image-Bench the model scores 60.28 overall — seventh of the 24 systems the chart ranks, behind GPT Image 2.5 Sunburst at 67.01 and behind Qwen's own hosted Qwen Image 3 Pro at 62.36. 1 3

Signal	Confirmed detail	What it means for you
What shipped	A 7B-parameter visual generation component with 32 single-stream DiT layers and native 2K output; ModelScope lists 16.22B parameters for the whole pipeline, the Qwen3-VL 8B text encoder included. 1 5	One prosumer GPU runs it with CPU offloading; the download is 33 GB.
What's new	Native RGBA throughout: generate a transparent image from a prompt, edit text inside a transparent layer, or lift a subject out of an ordinary photograph as a cut-out layer. 1	It absorbs December 2025's separate Qwen-Image-Layered model; compositing no longer needs a background-removal pass.
Editing	Up to 10 reference images; local edits marked with circles, painted strokes or a separate mask; portrait and product identity preserved. 1 6	Try-on, interiors and storyboards come from one model, not a chain.
The catch	The research license is non-commercial, and redistribution must carry the agreement and a Notice file. 4	Free to evaluate, not to ship; a product needs a commercial license first.

What the model actually is

Qwen-Image-2.1 is a single-stream diffusion transformer with block-causal attention: the system prefix and the edit instruction get a token-level causal mask, images get a chunk-level one. 2 The mixed granularity lets the model encode its reference images and instruction once and reuse that cached prefix on every denoising step — the source of Qwen's efficiency claim, which pays off most on edits carrying several input images. 2 A 64-channel RGBA autoencoder at 16× compression carries the alpha channel, and two fine-tuned Qwen3.5-VL 9B checkpoints ship beside the model to expand short prompts. 2

Where it stands, and where it stops

Qwen's Qwen-Image-Bench chart comparing overall score and parameter count across 24 image models — The chart is Qwen's own Qwen-Image-Bench, published with the release. Qwen-Image-2.1 lands at 60.28 with six systems above it, and the grey full-height columns are placeholders, as the chart's own footnote says, for models whose parameter counts are undisclosed — so the size comparison only holds against the open models. 1 3

Qwen's framing is a 7B model that "outperforms most closed-source models", which its own chart roughly supports for the systems ranked below it. 3 What the chart cannot settle is how the model does on a yardstick Qwen does not own, and no third-party evaluation exists yet: the release is hours old.

Before you build on it

Transparency is prompt-driven, not a switch: the model decides from the prompt whether to emit an ordinary image or one with an alpha channel, and the card says to ask for it outright. Everything else stays on default: 40 denoising steps, 2048×2048 output. 2 6

The license obligations outlive the download. Redistribution must carry the agreement and a Notice file; improving another model you distribute with these weights requires "Built with Qwen" in its documentation; and the agreement is governed by Chinese law, with disputes reserved to courts in Hangzhou. 4