Qwen3.8-27B-Uncensored-MLX (4 min read)

An abliterated version of Qwen3.8-27B with refusal responses removed has been released in MLX format. This model is optimized to run in Apple Silicon environments and offers 2/4/6/8-bit quantization options.

The model architecture adopts a hybrid structure combining Gated DeltaNet linear attention and full attention, while maintaining the vision tower, norms, and convolution layers in BF16 precision to preserve image understanding capabilities. In contrast, only the linear weights of the language model are quantized to improve efficiency.

Note: This model has had significant safety alignment removed and may respond to harmful or illegal requests. Therefore, it should only be used for legitimate research purposes such as AI safety research, interpretability analysis, and red team testing, and a custom safety and monitoring layer must be added when deploying.