What it is
Qwen-Image-Edit-2511 was released on December 23, 2025, as an update to Qwen-Image-Edit-2509. It edits images from plain-language instructions and can combine several input images, for example merging two people into one group shot. Alibaba's stated improvements are less image drift between edits, better character and multi-person consistency, popular community LoRAs built into the base model, and stronger geometric reasoning, such as drawing construction lines on a design.
The weights are released under Apache 2.0, with a 20.4-billion-parameter BF16 checkpoint that runs through Diffusers' QwenImageEditPlusPipeline. That licence is the main difference from Alibaba's newer Qwen-Image-2.1, which also edits but is free only for research and evaluation. Alibaba has published no newer open edit checkpoint since.
Our review
Qwen-Image-Edit-2511 is worth choosing for its licence, not its rank. On Artificial Analysis's image-editing board, listed as Qwen Image Edit Plus 2511, it scores 992 Elo and sits forty-ninth. The open models above it, Fibo 2.0 at 1089, Qwen-Image-2.1 at 1082 and HunyuanImage 3.0 Instruct at 1064, are roughly 70 to 100 points better.
Each of those comes with strings attached. Qwen-Image-2.1 needs a separate commercial licence from Alibaba, and HunyuanImage's licence excludes the EU, the UK and South Korea and wants eight 80GB GPUs. Qwen-Image-Edit-2511 carries no such conditions, which is why it still drew about 183,000 Hugging Face downloads in the past month, ten months after release.
The hosted price undercuts the case for Alibaba's own API. Artificial Analysis lists $60 per 1,000 edits there, twice the 2509 rate, while fal runs the same weights for $0.03 a megapixel, so a 1,024-pixel square edit costs about three cents. Alibaba's current Model Studio price list does not show the edit model at all; it leads with Qwen-Image-2.1 and 3.0, both ranked higher.
Self-hosting is where 2511 earns its keep. At 20.4 billion parameters the BF16 weights come to about 41GB before any offloading, so plan on an 80GB card or an FP8 build; Alibaba publishes no VRAM figure. DiffSynth-Studio claims it runs in 4GB with layer-by-layer offload at a heavy speed cost. LightX2V's distilled build cuts diffusion steps 25-fold, the difference between a batch job and an interactive editor.
Where it earns its place
An ML engineer at an e-commerce marketplace with one 80GB GPU
Batch-edit seller product photos (background swaps, removing props) without sending images to a third party
Serve the BF16 weights with vLLM-Omni or the LightX2V distilled build. The Apache licence means no per-image fees and no review of usage terms; the risk is quality on hard edits, where models 70 to 100 Elo higher do visibly better.
A product manager at a photo-app startup
Add a 'merge two friends into one photo' feature without a licence negotiation
Call fal at $0.03 a megapixel, so 100,000 one-megapixel edits a month cost about $3,000, then move to self-hosting at volume. Qwen-Image-2.1 would edit better but needs a commercial licence from Alibaba first.
An industrial designer at a hardware consultancy
Annotate concept renders with construction lines and variants
Use the geometric-reasoning improvements Alibaba highlights for 2511, prompting with the render and an instruction. Outputs are yours under Apache 2.0, though the results are Alibaba's own showcase, not independent tests.
Before you commit
- Setup
- pip install Diffusers and load QwenImageEditPlusPipeline from Hugging Face; Alibaba's sample uses 40 inference steps. ComfyUI, SGLang-Diffusion and vLLM-Omni also run it.
- Integrations
- Diffusers, DiffSynth-Studio (LoRA and full training), vLLM-Omni, SGLang-Diffusion, LightX2V (NVIDIA, Ascend and other accelerators), fal, Qwen Chat.
- Data and privacy
- Apache 2.0: commercial use, modification and redistribution allowed with no revenue cap or regional carve-outs. Self-hosted, nothing leaves your servers; Alibaba's image API purges task data and result URLs after 24 hours.
- Cost at scale
- On fal, 100,000 one-megapixel edits cost about $3,000 a month; Artificial Analysis lists $6,000 for the same on Alibaba's API. Self-hosted, the 20.4B-parameter BF16 weights are about 41GB, so one 80GB GPU such as an H100, or an FP8 build on a smaller card.
- Limits
- Forty-ninth on Artificial Analysis's editing board; no official VRAM figure published; Alibaba's improvement claims are self-reported showcases.
What's good
- Apache 2.0 with no revenue cap or regional exclusions
- Multi-image input for compositing and group shots
- fal hosts it at $0.03 per megapixel
- Day-one support in vLLM-Omni, SGLang-Diffusion and LightX2V
What's not
- Forty-ninth on Artificial Analysis's editing board at 992 Elo
- Alibaba's API price, $60 per 1,000 edits, is double 2509's
- 20.4B-parameter checkpoint needs a data-centre GPU at full precision
- No newer Apache-licensed edit model from Qwen since December 2025
Best for
Developers who need an open image editor they can deploy commercially without a licence negotiation.
Alternatives
Qwen-Image-Edit-2511: common questions
- Is Qwen-Image-Edit-2511 worth it?
- Qwen-Image-Edit-2511 is the open editor a legal team signs off without a meeting, and that is worth more than its forty-ninth place. We rate it 3.6 out of 5.
- How much does Qwen-Image-Edit-2511 cost?
- Free weights under Apache 2.0. Hosted: fal charges $0.03 per megapixel; Artificial Analysis lists $60 per 1,000 edits on Alibaba's API.
- What are the downsides of Qwen-Image-Edit-2511?
- Forty-ninth on Artificial Analysis's editing board at 992 Elo; alibaba's API price, $60 per 1,000 edits, is double 2509's; 20.4B-parameter checkpoint needs a data-centre GPU at full precision; no newer Apache-licensed edit model from Qwen since December 2025.
- Who is Qwen-Image-Edit-2511 best for?
- Developers who need an open image editor they can deploy commercially without a licence negotiation.
- How much does Qwen-Image-Edit-2511 cost for a team or in production?
- On fal, 100,000 one-megapixel edits cost about $3,000 a month; Artificial Analysis lists $6,000 for the same on Alibaba's API. Self-hosted, the 20.4B-parameter BF16 weights are about 41GB, so one 80GB GPU such as an H100, or an FP8 build on a smaller card.
- Is Qwen-Image-Edit-2511 safe for company data?
- Apache 2.0: commercial use, modification and redistribution allowed with no revenue cap or regional carve-outs. Self-hosted, nothing leaves your servers; Alibaba's image API purges task data and result URLs after 24 hours.
- What are the best Qwen-Image-Edit-2511 alternatives?
- The closest alternatives we have reviewed are Qwen-Image-2.1, HunyuanImage 3.0 Instruct, FLUX 3 Image Klein and Nano Banana 2.1.
Answers checked Oct 22, 2026.