Qwen3.5-27B abliterated fine-tune ships on HuggingFace for unrestricted multimodal inference

DavidAU released Qwen3.5-27B-Polar-Rev1-Uncensored-Heretic on HuggingFace on July 10, an abliterated fine-tune of Alibaba's Qwen3.5-27B-Polar-Rev1 multimodal model. The weights are tagged "uncensored," "abliterated," and "heretic," signaling removal of safety alignment layers that typically filter outputs in commercial releases. The model handles image-text-to-text tasks and ships as safetensors for local inference, compatible with the transformers library and runnable on consumer GPUs with sufficient VRAM for 27-billion-parameter models in half-precision or quantized formats.

Abliteration refers to the surgical removal of safety refusal mechanisms embedded during reinforcement learning from human feedback—a technique that has gained traction among practitioners running models locally without API-layer content filters. Multimodal variants like Polar-Rev1 extend text-only Qwen base models to handle image inputs alongside text prompts, a capability that has driven adoption in vision-language tasks ranging from OCR to image captioning to visual question answering. The checkpoint is marked not-for-all-audiences on HuggingFace, a designation the platform applies to models that lack content filtering or carry adult-oriented fine-tuning. DavidAU has released similar uncensored variants across the Qwen and Llama families, building a catalog of unrestricted checkpoints for users who need full control over model behavior. At publication, the model card showed zero downloads and two likes.