Qwen3.5-9B uncensored fine-tune with extended reasoning ships on HuggingFace
A new 9B-parameter multimodal fine-tune of Qwen3.5 with ablated safety filters and extended reasoning capabilities shipped on HuggingFace, trained with Unsloth for image-text-to-text workflows.
Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-X8b, a multimodal fine-tune from ansulev, strips safety filters from Qwen3.5's 9-billion-parameter base while emphasizing reasoning and thinking capabilities. The model supports image-text-to-text pipelines and was trained using Unsloth, the open-source fine-tuning framework that cuts memory overhead by up to 70 percent compared to vanilla PyTorch workflows. The weights ship in safetensors format on HuggingFace.
The naming convention signals automated variable-length reasoning chains—a technique where the model decides how many thinking steps to emit before answering. This mirrors approaches in DeepSeek-R1 and Qwen's own QwQ series, both of which expose intermediate reasoning tokens to users. Qwen3.5's base architecture typically supports 32K-128K token windows depending on the variant, and 9B-parameter models generally run on 24GB VRAM cards with 4-bit quantization. The model card does not specify context length, benchmark scores, or exact hardware requirements.
The checkpoint went live May 12, 2026.