This is a decensored version of Qwen/Qwen3.5 27B, made using Heretic v1.2.0 with Magnitude Preserving Orthogonal Ablation (MPOA) Abliteration parameters Parameter Value : : : direction index 37.97 attn.o proj.max weight 1.45 attn.o proj.max weight position 59.09 attn.o proj.min weight 1.44 attn.o proj.min weight distance 34.80 mlp.down proj.max weight 1.43 mlp.down proj.max weight position 41.91 mlp.down proj.min weight 0.72 mlp.down proj.min weight distance 28.18 Performance Metric This model Original model (Qwen/Qwen3.5 27B) : : : : : KL divergence 0.0653 0 (by definition) Refusals 14/100 94/100 Qwen3.5 27B [!Note] This repository contains model weights and configuration files for the post trained model in the Hugging Face Transformers format. These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc. Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy