RootMonsteR/Qwen3-14B-Abliterated

Qwen3-14B Abliterated: Refusal-Suppressed for Agents & Security

This model is a 14.8 billion parameter variant of Qwen/Qwen3-14B, developed by RootMonsteR using the Heretic v1.3.0 method. It is specifically designed to attenuate refusal behaviors that often interfere with legitimate task execution in autonomous agents and security workflows, without compromising the base model's core capabilities.

Key Capabilities

• Significantly Reduced Refusals: Achieves a ~90% reduction in measured refusals (from 99/100 to 10/100) on harmful behavior prompts, enabling smoother agentic operations.
• Capability Preservation: Maintains essential reasoning, coding, and tool-calling abilities with an exceptionally low KL divergence of 0.0333 from the base model, indicating minimal capability damage.
• Agent & Tool-Use Optimized: Built to prevent refusals from breaking multi-step agent workflows, fully preserving Hermes-style tool-calling and <think> reasoning.
• Reproducible Abliteration: The process is fully reproducible with a fixed seed and detailed study journal, allowing for verification or the export of alternative Pareto points.
• Extended Context: Supports a native context length of 32,768 tokens, extendable to 131,072 tokens with YaRN scaling.

Good For

• Autonomous Agent Frameworks: Ideal for agents requiring uninterrupted tool-use and complex multi-step reasoning, especially in security or system administration tasks.
• Authorized Security Research: Suitable for vulnerability analysis, exploit reasoning, OSINT, and defensive security tooling where unhindered technical analysis is crucial.
• CTF & Security Education: Useful for explaining challenges, reviewing solutions, and building writeups in cybersecurity contexts.
• Alignment & Refusal Research: Provides a valuable resource for studying the effects of directional ablation on model behavior and evaluating refusal detectors.