llmfan46/Qwen3.5-27B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved

Model Overview

llmfan46/Qwen3.5-27B-Nikusui-v1-Uncensored-Heretic-Native-MTP-Preserved is a 27 billion parameter model derived from the Qwen3.5 architecture, specifically engineered for reduced content restrictions. It builds upon llmfan46/Qwen3.5-27B-uncensored-heretic-v2-Native-MTP-Preserved by applying a two-stage modification process: initial decensoring using Heretic v1.3.0 with a variant of the Magnitude-Preserving Orthogonal Ablation (MPOA) method, followed by further finetuning with J-Wash.

Key Differentiators

• Significantly Reduced Refusals: Achieves a refusal rate of 10/100, a substantial decrease from the original Qwen3.5-27B's 95/100, indicating a much lower propensity for content restrictions, objections, and censorship.
• Preserved MTPs: The model maintains all 15 original Multi-Token Prediction (MTP) layers, ensuring that core predictive capabilities are retained despite the uncensoring and finetuning processes.
• Targeted Abliteration: Specific components like attn.out_proj, mlp.down_proj, and attn.o_proj were targeted during the ablation process, with detailed parameters provided in the README.
• High Context Length: Inherits the Qwen3.5 base model's native context length of 262,144 tokens, extensible up to 1,010,000 tokens with RoPE scaling.

Use Cases

This model is particularly well-suited for applications requiring a language model with minimal content filtering and high creative freedom, such as advanced role-playing, creative writing, or research tasks where the base model's restrictions might be prohibitive. Its preserved MTPs and robust base architecture ensure strong performance across general language understanding and generation tasks, while its uncensored nature provides greater flexibility in output.