Alibaba's Qwen Team open-sourced Qwen-Image-2.1 on September 20, 2026, releasing a unified text-to-image generation and editing model whose visual generation component uses 7 billion parameters across 32 single-stream diffusion transformer layers3,5,6. Weights landed simultaneously on Hugging Face, ModelScope, and GitHub, with day-zero integration in ComfyUI, Diffusers, vLLM-Omni, and SGLang9.
From 20B to 7B
The original Qwen-Image shipped in August 2025 as a roughly 20-billion-parameter model under Apache 2.0, with editing handled by a separate checkpoint. Qwen-Image-2.1 merges both jobs into a single checkpoint at about a third of the parameter count. The 7B figure covers the diffusion transformer only; the full pipeline also loads a Qwen3-VL 8B text encoder and a 64-channel RGBA variational autoencoder with 16× spatial compression, bringing the core pipeline to roughly 15B parameters.
Alibaba Group Holding positions the release as a compact alternative targeting strong quality-per-compute on consumer GPUs such as an RTX 3090-class card. A test documented during Diffusers integration measured peak memory of 56.5 GiB for a single 2048×2048 image on one Nvidia H100 in bfloat16, completing in 32.7 seconds.
What the single checkpoint covers
Qwen-Image-2.1 natively generates and edits transparent RGBA images, letting users isolate subjects or manipulate text on transparent layers without a separate matting model4. The model accepts up to ten reference images in one pass for identity-preserving group composites, virtual try-on, and room restyling. Local edits are guided by circles, painted marks, or external masks so that global restyles do not overwrite untouched regions. Native 2K output defaults to 2048×2048 and supports seven aspect ratios up to 2752×1536.
Architecture changes including mixed-granularity attention and prefix KV-cache reuse compute text and reference images once at the first step and reuse the cache for every remaining step, intended to keep multi-reference edits interactive.
Benchmark claims and licensing
On Qwen-Image-Bench, the team's own evaluation suite, Qwen-Image-2.1 scored 60.28, above Google's Nano Banana 2.0 at 59.82 and FLUX 2 Max at 55.33, though below GPT Image 2.5 Sunburst at 67.01. Independent benchmarks are still pending. On GenAI Showdown, a separate evaluation, Qwen-Image-2.1 scored 7 out of 15, compared with 4 for Qwen-Image 1.0 and 8 for Ideogram 4.
The release ships under the Qwen Research License, which bars commercial use. Businesses must apply separately to Hangzhou Tongyi Laboratory Technology Co. for a commercial grant. That marks a shift from the original Qwen-Image's Apache 2.0 terms.
Qwen Team also released two prompt-rewriting models fine-tuned from Qwen3.5-VL 9B checkpoints, which expand short prompts into detailed ones and can select an aspect ratio, adding roughly 18.8 GB each. Serving-side tooling includes FP8 quantization and CUDA Graph decode via vLLM-Omni, Cache-DiT and multi-GPU parallelism via SGLang, and native ComfyUI nodes. AMD Radeon GPUs are supported via ROCm.
ANALYSIS The parameter compression from 20B to 7B while consolidating generation and editing into one checkpoint lowers the hardware floor for local experimentation, but the switch from Apache 2.0 to a restrictive research license narrows the commercial path relative to the original release.