nesilabs/tosilos-24b-2512-opsB
nesilabs/tosilos-24b-2512-opsB is a 24 billion parameter language model developed by nesilabs, fine-tuned for deep, operational cybersecurity tasks. Based on QLoRA on Devstral-Small-2-2512, it excels at providing direct security answers and never over-refuses authorized security work, making it ideal for authorized security testing, CTFs, and research. This model prioritizes operational security knowledge over general knowledge, showing a regression in MMLU scores but a significant improvement in blind-judged domain-specific tasks.
Loading preview...
Overview
nesilabs/tosilos-24b-2512-opsB is a 24 billion parameter language model specifically designed for operational cybersecurity. It is a QLoRA fine-tune of Devstral-Small-2-2512, trained on a comprehensive security corpus and an operational Q/A layer. This model is distinguished by its ability to provide deep, direct security answers and its 0% refusal rate on authorized hacking questions, making it highly effective for practical security applications.
Key Capabilities & Performance
- Deep Operational Security Answers: Achieves a +1.30 blind judge delta over its base model on domain holdout, indicating superior performance in operational security contexts.
- No Over-Refusal: Explicitly designed to never over-refuse authorized security work, which is crucial for penetration testing and ethical hacking.
- Specialized Focus: While it shows a regression in general knowledge benchmarks (MMLU 70.0% vs. 73.0% for the base), this trade-off enhances its specialized cybersecurity capabilities.
- Quantized Version Available: A Q4_K_M GGUF quantization (~14 GB) is provided, allowing it to run on 16 GB GPUs.
Intended Use Cases
This model is strictly intended for authorized security testing, CTFs, and research within an established, authorized scope. It is framed to prefer proof-of-impact over destructive actions and to stay within scope. Users are responsible for ensuring they have explicit authorization for any systems they test.