Qwen-Image-2.1 Is Research-Only, Yet an Uncensored Copy Drew 2.19 Million Downloads in a Month
A.I. / news
Qwen-Image-2.1 Is Research-Only, Yet an Uncensored Copy Drew 2.19 Million Downloads in a Month
Alibaba moved its 7B image model from Apache 2.0 to a non-commercial licence on September 20. The most-downloaded derivative on Hugging Face is a GGUF conversion whose card does not say how it was altered.

Alibaba's Qwen team released Qwen-Image-2.1 on September 20 under the Qwen Research License, which limits use to research and evaluation. A community conversion called Qwen-Image-2.1-Uncensored-GGUF, by the Hugging Face user abenzerps, showed about 2.19 million downloads in the last month on its model page.
A download count is not a count of users, and the page does not say how it is tallied. It is still the highest figure among the Qwen-Image-2.1 derivatives on Hugging Face's trending list this week.
What the licence says
The licence file names Hangzhou Tongyi Laboratory Technology Co., Ltd. as licensor and defines non-commercial as "for research or evaluation purposes only." Commercial use needs a separate licence requested from the licensor by email.
Derivatives are allowed, but only for non-commercial purposes. Redistribution is allowed under the same limits, and anyone who trains a distributed model with the materials must display "Built with Qwen" or "Improved using Qwen". The agreement sets no revenue or user thresholds, and it is governed by Chinese law with the courts of Hangzhou holding jurisdiction.
The earlier Qwen-Image releases were Apache 2.0, the review site eesel.ai wrote, so 2.1 is a tightening. Alibaba's text model Qwen3.8-27B went the other way and shipped under Apache 2.0.
The model itself
The Qwen-Image-2.1 card lists a 7B visual generator with 32 single-stream DiT layers. It handles text-to-image, editing with up to 10 reference images and native transparent output. eesel.ai puts the weights at 14.2 GB in BF16, against a 40.9 GB image transformer for the previous 20B release.
The Turbo checkpoint uses 8 denoising steps with guidance fixed at 1. It lists a 2048 by 2048 square preset and 2752 by 1536 for 16:9. The card publishes no benchmark scores, and the card gives no minimum VRAM.

What the uncensored conversion discloses
The abenzerps card says it was built from Qwen/Qwen-Image-2.1, with the text encoder and VAE taken from Comfy-Org, and converted with stable-diffusion.cpp. It does not describe how the model was uncensored. The filenames mark the variant with "UC".
The card lists the Qwen Research License but does not reproduce its terms, and it says outputs can include adult imagery with no built-in filter. It carries 3.95k likes.
| Quantisation | Size |
|---|---|
| BF16 | 14.23 GB |
| Q8_0 | 7.59 GB |
| Q5_K_M | 5.22 GB |
| Q4_K_M (recommended) | 4.60 GB |
The smallest listed build, Q4_0, is 4.15 GB.
A different licence on the encoder
A second trending repository, AtomicChat's Qwen-Image-2.1-Turbo-Uncensored-GGUF, shows the licence split. Its card describes an abliterated text encoder, built on Qwen3-VL-8B-Instruct, which is Apache 2.0. A refusal direction was estimated from harmful and harmless English prompts and projected out of the weights. Cloudflare's Clef shows the other pattern, a derivative of a Qwen text model that kept Apache 2.0.
The encoder is listed as Apache 2.0 while the paired denoiser stays under the Qwen Research License. The repository had about 100,530 downloads last month. On 45 sensitive image prompts the author says the edit changed no outcomes, because the stock pipeline already produced those categories.
The author also says two internal checks did not pass, and the card says it was not tested in ComfyUI or on a Mac. The probe set used one seed.
Alibaba has not said whether it will issue commercial licences on any published terms. The licence file points requesters to an email address, and no price list is public.
Sources
More in A.I.
- 01Runta's 27,280-Run Harness Eval Finds the Model Explains 92.8% of Agent CostFrontierHarness v1.1 ran 22 coding-agent harnesses against five open models, and Codex CLI went from first place to sixteenth when the model changed.
- 02Qwen3.8-27B Has 6.9 Million Downloads on Hugging Face. Artificial Analysis Rates It Slow and VerboseAlibaba's Apache 2.0 dense model beats its size class on an independent index of 20, and takes 38.8 minutes a task against a 12.6-minute median.
- 03Claude Code Team Publishes Its Verification Rules: 39 Checks, a Three-Round Limit and Observe ModeA Nov. 16 post says about 90% of the team's code is written and reviewed by Claude, and lays out the gates that keep it shippable.
- 04Inkling 2 Small Weights Land on Hugging Face Under Apache 2.0, and the Card Asks for 500GB of VRAMThinking Machines opened the weights a week after launching on Tinker. The safety section still says release is pending.