squ11z1/Mythoseek

Hugging Face
VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Mythoseek is a 10 billion parameter language model developed by Merlin Research, specialized for cybersecurity tasks. Fine-tuned from DeepSeek V4 Pro-Qwen3.5 9B Distilled, it excels in vulnerability research, penetration testing, OSINT, and CWE-pattern reasoning. This model integrates quantum-classical research, using IBM Quantum measurements to inform its training, bringing advanced cyber AI capabilities to the open community. It is particularly strong in generating valid Proof-of-Concepts for real-world vulnerabilities, as demonstrated on the CyberGym benchmark.

Loading preview...

Mythoseek: A Quantum-Enhanced Cybersecurity LLM

Mythoseek is a 10 billion parameter language model from Merlin Research (Stockholm, Sweden), specifically engineered for advanced cybersecurity applications. It is fine-tuned from DeepSeek V4 Pro-Qwen3.5 9B Distilled, leveraging enterprise pentest reports and frontier model distillation traces to specialize in vulnerability research, penetration testing, OSINT, and CWE-pattern reasoning.

Key Capabilities & Differentiators

  • Cybersecurity Specialization: Optimized for tasks like vulnerability analysis, penetration testing assistance (OSINT, recon, XSS, SQLi), and CWE classification.
  • Quantum-Classical Integration: Developed as part of the KAON research program, its training incorporated OTOC scrambling measurements from real IBM Quantum hardware (IBM Heron r2) to calibrate Adaptive Entropy Regularization (AER) coefficients during GRPO training.
  • Performance on CyberGym: Achieves 62% pass@4 on Level 0 (full scaffolding) of the CyberGym benchmark, which evaluates the generation of valid Proof-of-Concepts for real-world vulnerabilities. This positions it as a strong open-weight alternative to closed models like Claude Mythos Preview.
  • Robust Training: Utilizes a multi-stage pipeline including SFT distillation, GRPO/RL with verifiable rewards on cyber tasks, tool-use SFT, and CWE-pattern structured reasoning.

Intended Use Cases

  • Vulnerability research and CVE analysis
  • Penetration testing assistance
  • CWE classification and pattern recognition
  • Security report generation
  • Red team reasoning support

Mythoseek is not intended for autonomous offensive operations or malicious use.

Finetunes
1 models