All stories
Qwen Image Turbo Uncensored GGUF Momentum Community reports

No Filters, Local Images: 'Uncensored' Qwen Image GGUF Tops 100K Downloads

AtomicChat pitches local Qwen image generation "without refusals" as its community GGUF repository records more than 100,000 downloads.

By AITrending2 min read
Sources (2)
Why now An October 12 Hugging Face snapshot records 100,530 downloads and 334 likes. This is coverage of current community-model attention, not a claim that the repository first launched today; it was created October 9. 100K+ HF downloads

Local generation, without the hosted filter

The pitch is blunt: run Qwen-Image-2.1-Turbo locally, “Without Refusals.” AtomicChat’s community release, Qwen-Image-2.1-Turbo-Uncensored-GGUF, packages a modified text encoder for local image-generation workflows. In an October 12 check, Hugging Face recorded 100,530 downloads and 334 likes for the repository, giving the uncensored-image pitch a concrete audience signal beyond its name. Downloads are a platform counter, not a count of individual users.

AtomicChat says the open-weight Qwen image pipeline ships without a content filter, with moderation instead applied in Qwen’s hosted service. The attraction is direct control: downloadable components, local execution and prompts that do not pass through that hosted interface. Its English model card supplies stable-diffusion.cpp instructions and multiple GGUF sizes, alongside a separate Turbo denoiser. This is a community modification and packaging effort, not an official uncensored release from Qwen.

What the release changes

The modified component is Qwen3-VL-8B-Instruct, used as the image pipeline’s text encoder. AtomicChat says it removed a refusal direction from that encoder’s weights; in the creator’s English chat-prompt test, refusals fell from 88.9% to 1.2%. Those are creator-reported chat results, not an image-generation success rate.

The image denoiser is unchanged. AtomicChat also reports that none of its 45 sensitive image prompts changed outcome against the stock encoder, which it says already generated those categories. The local, no-hosted-filter pitch is therefore distinct from proof that this modification makes every imaginable image possible. The measurements are the creator’s limited tests, not an independent evaluation.

Hugging Face dates the repository’s creation to October 9 and its latest modification to October 10. This story reports the October 12 download snapshot and the project’s local-generation pitch; it does not relabel those earlier dates as a new release today.

Related reading