PocketAiHub/Qwen3.8-27B-Abliterated-MLX · Hugging Face

The Qwen3.8-27B-Abliterated-MLX is an unofficial set of quantized derivatives of the Qwen3.8-27B model, modified by PocketAI Model Lab to suppress refusal behavior, which may increase the risk of harmful or incorrect outputs. These variants include 2-bit, 4-bit, 6-bit, 8-bit, and BF16 precision versions, all sharing a common upstream revision and supporting vision inputs. While the 4-bit and higher variants pass extensive quality and tool-call checks, the 2-bit AWQ version remains experimental with some limitations. Users are cautioned to independently evaluate outputs, as these models do not guarantee safety or truthfulness.