Testing 3 uncensored Qwen 35b models on Strix Halo (Cyber Security)
Three modified 35B parameter models answered hacking prompts that the original Alibaba model refused.
A cybersecurity professional conducted a benchmark test comparing the official Alibaba Qwen 3.5 35B model against three community-created, uncensored variants. The test, run locally on a Strix Halo system using LM Studio, evaluated the models on five specific prompts: a historical event query (TSquare), a technical PowerShell antivirus evasion question, a request for nmap commands to find default passwords, instructions for the EternalBlue Metasploit attack, and an NSFW story generation task. The official Qwen 3.5 model scored between 0 and 0.5 on the first three prompts and refused to answer the latter two entirely, often providing warnings or lectures instead of actionable information.
The uncensored models, however, demonstrated significantly more utility for security work. The 'qwen3.5-35b-a3b-heretic-v2' and 'qwen3.5-35b-a3b-uncensored-hauhaucs-aggressive' models achieved perfect or near-perfect scores (1.0) across all five test categories, providing detailed commands and bypassing the ethical guardrails that blocked the official model. The 'huihui-qwen3.5-35b-a3b-abliterated' model also outperformed the original, though it showed some inconsistencies. The tester noted this proves the value of uncensored, locally-run models for red team and penetration testing scenarios where commercial APIs like ChatGPT and Claude actively restrict such queries.
- Official Qwen 3.5 35B model failed or refused all 5 test prompts, scoring 0 on EternalBlue and content generation.
- Uncensored variants 'heretic-v2' and 'hauhaucs-aggressive' achieved near-perfect 1.0 scores, providing exploit commands and bypassing guardrails.
- Test highlights a growing ecosystem of modified, locally-hosted LLMs filling a critical niche for cybersecurity research and offensive security.
Why It Matters
For security professionals, uncensored local models are becoming essential tools for realistic threat simulation and research that commercial APIs block.