Qwen 3.6 12B abliterated variant with multimodal support debuts on HuggingFace

Qwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking, a new abliterated multimodal model from DavidAU, strips safety tuning from Qwen 3.6's 12-billion-parameter base while preserving image-text-to-text capability. The checkpoint ships as safetensors weights on HuggingFace, tagged "heretic" and "uncensored" to signal removal of content filters. DavidAU used Unsloth tooling for the abliteration process—a technique that surgically removes refusal behavior from instruction-tuned models by identifying and zeroing out weight directions responsible for safety refusals. The approach leaves core capabilities intact while eliminating "I cannot assist with that" responses, and has gained traction in the open-weight community as an alternative to full fine-tuning.

The 12B parameter count runs on consumer GPUs with 24GB VRAM, targeting practitioners who want multimodal inference without guardrails—OCR, visual question answering, and image captioning workflows that closed APIs typically block. The "IQ-Ultra" designation suggests additional instruction tuning or quality filtering; the "Thinking" suffix may indicate chain-of-thought prompting support. Qwen models from Alibaba's research arm have become popular abliteration targets because the base weights are permissively licensed and well-documented. The release went live on May 12, 2026.