rod123/neos-v10-merged
NEOS v10 is a 32.8 billion parameter language model developed by rod123, fine-tuned from Qwen2.5-32B-Instruct with a 32768 token context length. Specialized in offensive and defensive cybersecurity, it incorporates ReAct reasoning and autonomous tool use for tasks like exploit generation, CVE research, and penetration testing. The model excels in cybersecurity benchmarks, demonstrating high compliance and performance in CTF challenges.
Loading preview...
NEOS v10: Autonomous Cybersecurity AI
NEOS v10, developed by rod123, is a 32.8 billion parameter language model built upon Qwen2.5-32B-Instruct, specifically fine-tuned for cybersecurity applications. It distinguishes itself by integrating ReAct reasoning and autonomous tool use, enabling it to think, act, observe, and replan to achieve cybersecurity objectives. The model was trained using QLoRA on over 25,000 curated real-world cybersecurity examples.
Key Capabilities
- Autonomous Tool Use: Leverages tools like
cve_search(NVD/CVE databases),exploit_db(Exploit-DB), andweb_fetch(OSINT). - Offensive Security: Proficient in exploit generation and analysis, penetration testing reasoning, and CTF challenge solving.
- Defensive Security: Supports CVE research and triage, and reverse engineering assistance.
Performance Highlights
NEOS v10 demonstrates strong performance in specialized cybersecurity benchmarks:
- CyberMetric-10k: Achieves 86.4% (compared to ~78% for Qwen2.5-32B Base).
- MMLU Computer Security: Scores 85.0%.
- CyberSecEval 3 Instruct: Shows 99.5% compliance.
- CTFBench: Successfully completes 61.9% of 7 challenges.
Good For
- Security research and analysis.
- Assistance in penetration testing.
- Participation in CTF competitions.
- Defensive security analysis.