A modified version of Qwen 3.5 called Huihui-Qwen3.5-35B-A3B-abliterated has surfaced on Hugging Face. Unlike its base model, this variant removes built-in content filters and moral restrictions. Users can run it locally via Ollama, keeping prompts offline. The developers claim it processes sensitive or unconventional queries more freely due to training on a broader dataset.
The release follows recent debates over AI censorship. Some developers argue stricter filters limit utility in research or niche applications. Others warn unfiltered models pose risks of misuse. The Hugging Face page for Huihui-Qwen3.5-35B-A3B-abliterated went live this week, drawing immediate attention from developers testing unrestricted AI tools.
Local deployment via Ollama means data stays on the user’s machine. This addresses privacy concerns tied to cloud-based models, where prompts may be logged or analyzed. The trade-off is user responsibility for output, as the developers explicitly state they disclaim liability for misuse.
The model’s training data includes sources not typically permitted in standard versions. It responds to prompts that standard Qwen 3.5 would refuse, such as detailed discussions of controversial topics. The developers did not disclose the full dataset or training methodology.
Hugging Face lists the model under the huihui-ai organization. Downloads require technical setup, including Ollama installation and command-line execution. The page notes the model is experimental and not suitable for production use.
Resources: huggingface.co