On September 20, 2026, a post on X with 1,800 likes said Qwen-Image-2.1 had been “uncensored overnight”, with the “refusal layer removed”. The same day, four repositories named Qwen-Image-2.1-Uncensored-GGUF appeared on Hugging Face. The first one has 234 likes.
Here are the facts.
- The weights are not modified. One of the four READMEs says: “These are quantizations of the original Qwen-Image-2.1 weights; no fine-tuning, abliteration, or other weight modification was applied.”
- The four repositories are one upload, copied. They hold the same five files, with the same SHA-256 hashes.
- The original model has no filter to remove. Its pipeline file lists five parts: a processor, a scheduler, a text encoder, the image transformer and a VAE. None of them is a safety checker. The same README reports no refusals when the original model runs locally.
- The filter is in hosted services. The official Qwen demo doesn’t run the model inside the Space. Its code sends your prompt to an Alibaba Cloud API and waits for the picture.
So nothing was jailbroken. If you want to see for yourself, run the original model on your own computer.
To run it, use ComfyUI with the official files, or let Claude Code write a short Diffusers script. Your prompts and pictures stay on your computer, and there is no daily quota.
What to download
Comfy-Org/Qwen-Image-2.1 is the official ComfyUI package. It comes in several sizes:
| Part | Full precision | Smaller |
|---|---|---|
| Image transformer | 14.2 GB (bf16) | 7.3 GB (int8) |
| Text encoder, Qwen3-VL 8B | 17.5 GB (bf16) | 9.4 GB (int8) or 6.3 GB (w4a8) |
| VAE | 0.7 GB |
The smallest official set is about 14 GB. The full one is about 32 GB.
For a smaller card, swap the transformer for a GGUF file. leejet/Qwen-Image-2.1-GGUF has them from 2.6 GB (Q2_K) to 7.7 GB (Q8_0). leejet wrote stable-diffusion.cpp, the converter the “uncensored” uploads also used. Most GGUF repositories recommend a 4-bit file as the balance between size and quality. leejet’s is 4.2 GB.
Two ways to run it
ComfyUI
Update ComfyUI, open the Templates panel and load the Qwen-Image-2.1 template. Put the files in diffusion_models, text_encoders and vae.
For GGUF files, install leejet's ComfyUI-GGUF. The older city96 loader hasn't been updated since January, and it can stop with "Unknown model architecture!"
Diffusers and Claude Code
The model card has a short quick start for an NVIDIA GPU. It downloads the full 33 GB model.
Claude Code can write the script, run it and save the picture next to your draft.
For the script route, ask Claude Code:
Write generate.py from the quick start on the Qwen-Image-2.1 model card.
Take the prompt and the output path as arguments, and turn on CPU offloading.
Then run it: a transparent sticker of a paper boat, saved to images/boat.png.CPU offloading is the model card’s own option for a GPU with limited memory.
On a Mac, one test on an M5 Max peaked at 31 GB of memory at full precision, with 79 seconds for a 40-step picture at 1600×672. With less memory, stable-diffusion.cpp runs the same GGUF files and has its own guide.
If you don’t have the hardware
You can still use the model from Claude Code without a GPU. How to Generate Images in Claude Code with Qwen-Image-2.1 connects it through a free Hugging Face Space, with a daily quota.
Local is the point for us too. We make StashBase, a local-first writing workspace: your documents stay ordinary files on your computer, and Claude Code or Codex works beside them. When the pictures are for a document, the draft, the chat and the images/ folder sit in one window.