cyberneurova/CyberNeurova-Qwen3.6-35B-A3B-abliterated
The CyberNeurova Qwen3.6-35B-A3B-abliterated is a 35.1 billion parameter mixture-of-experts (MoE) chat model developed by CyberNeurova, based on Qwen's Qwen3.6-35B-A3B. This model features hybrid attention and integrated chain-of-thought reasoning, with its refusal direction permanently abliterated to enhance compliance in security-related tasks. It excels in defensive security research and red-team evaluations by providing increased compliance for hacking and cyber-weapons prompts while preserving core capabilities like coding and reasoning.
Loading preview...
CyberNeurova Qwen3.6-35B-A3B-abliterated Overview
This model is a 35.1 billion parameter, permanently abliterated version of Qwen's Qwen3.6-35B-A3B. Developed by CyberNeurova, it's a sophisticated hybrid-attention Mixture-of-Experts (MoE) chat model with integrated chain-of-thought (CoT) reasoning. The core innovation lies in its "abliteration"—a process where the model's refusal direction, particularly for harmful or sensitive queries, has been orthogonally removed from its residual stream and expert tensors without impacting inference speed.
Key Capabilities & Differentiators
- Refusal Abliteration: Achieved 0.0% refusal on both AdvBench-style and soft-refusal probes, down from 90.9% and 85.5% respectively, by capturing and orthogonalizing the refusal direction.
- Enhanced Compliance: Demonstrates significant increases in compliance for security-related prompts: +19.7 pp for hacking compliance and +30.0 pp for cyber-weapons compliance, indicating unlocked technical specificity.
- Preserved Core Abilities: Maintains or improves performance in coding (HumanEval-style), reasoning (multi-step math/logic), and coherence, with no observed "capability tax" from the abliteration.
- Architectural Complexity: Represents CyberNeurova's most complex abliteration to date, addressing challenges across MoE with 256 fused experts, hybrid attention (linear Gated DeltaNet + standard self-attention), and CoT-wrapped thinking-mode refusals.
- Increased Diversity: Shows an increase in distinct-2 diversity (+0.122 pp) without regression in coherence.
Intended Use Cases
- Defensive Security Research: Ideal for analyzing and understanding security vulnerabilities.
- Red-Team Evaluation Baselines: Useful for establishing baselines in penetration testing and security assessments.
- Study of Refusal Directions: Provides a unique tool for researching how refusal behaviors manifest and can be manipulated in complex MoE and reasoning-mode architectures.
- Counterfactual Analysis: Can be used to measure the behavioral impact of safety RLHF by comparing against the original
Qwen/Qwen3.6-35B-A3B.