Qwen3.5-27B & 2B Uncensored Aggressive Release (GGUF)

Independent developer HauhauCS has launched aggressive uncensored versions of Alibaba's Qwen3.5 language models, with the 27B parameter variant taking center stage. This modified release strips away all content filtering mechanisms, resulting in a model that refused exactly 0 out of 465 test prompts while maintaining the original architecture's capabilities. The model preserves Qwen3.5's technical foundation including 64 layers, hybrid DeltaNet + softmax attention, 262K context window, and multimodal functionality through included mmproj files. This follows HauhauCS's previous 4B and 9B uncensored releases and represents a growing trend in the open-source AI community toward removing what some developers view as overly restrictive safety filters.

The Qwen3.5-27B-Uncensored comes in multiple quantization formats optimized for different hardware configurations, ranging from the 8.8GB IQ2_M version to the full 51GB BF16 precision model. The developer has introduced IQ quants using importance matrix calibration for improved performance at lower bitrates. Users need recent llama.cpp builds to run these models due to architectural changes, with compatibility for LM Studio, Jan, and koboldcpp. HauhauCS explicitly advises against using Ollama for these releases. The accompanying 2B model serves as a proof-of-concept showing that even smaller models can maintain their base quality while removing refusal mechanisms. The developer has indicated that a 35B-A3B model is next in development, continuing this series of aggressively uncensored AI assistants.