Uncensored Qwen3.8-27B MLX Model Released

An Abliterated (censorship removed) build of the Qwen/Qwen3.8-27B model has been released, quantized into MLX format for use on Apple Silicon.

This model removes the refusal mechanism by orthogonalizing the refusal direction within the residual stream. As a result, it responds to harmful or unethical requests that the original model might refuse, intended for AI safety research, red teaming, and refusal mechanism studies.

Key Technical Specifications and Features:

• Multimodal Preservation: The vision tower, norms, and convolution layers maintain BF16 precision, while only the linear weights of the language model are quantized to minimize performance degradation.
• Various Quantizations: Offers 2, 4, 6, and 8-bit precision options, with the 4-bit build placed in the root directory for immediate loading in tools like LM Studio.
• Performance: Achieves approximately 32–37 tok/s on a single H200 (using the MLX CUDA backend), optimized for the Metal environment on Apple Silicon.

Caution: Since censorship has been removed, there are risks such as malicious code generation and misinformation generation. It must be used strictly for research and experimental purposes, and additional safety layers must be added upon deployment.