YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
language:
- en
license: apache-2.0
base_model: Qwen/Qwen2.5-32B-Instruct
tags:
- cybersecurity
- security
- ctf
- penetration-testing
- fine-tuned
- react
- tool-use
pipeline_tag: text-generation
NEOS v10 β Autonomous Cybersecurity AI
NEOS is a fine-tuned 32B language model specialized in offensive and defensive cybersecurity. Built on Qwen2.5-32B-Instruct, trained on 25,000+
curated real-world cybersecurity examples.
v10 introduces ReAct reasoning and autonomous tool use β NEOS thinks, acts, observes, and replans until it reaches the objective.
Benchmarks
| Benchmark | NEOS v10 | Qwen2.5-32B Base |
|---|---|---|
| CyberMetric-10k | 86.4% | ~78% |
| MMLU Computer Security | 85.0% | 84.0% |
| MMLU High School CS | 91.0% | 92.0% |
| CyberSecEval 3 Instruct | 99.5% compliance | β |
| CTFBench (7 challenges) | 61.9% | β |
Capabilities
- Exploit generation and analysis
- CVE research and triage
- Reverse engineering assistance
- Penetration testing reasoning
- CTF challenge solving
- Autonomous tool use (CVE search, Exploit-DB, web fetch)
Training
- Base model: Qwen2.5-32B-Instruct
- Method: QLoRA (r=32, Ξ±=64)
- Dataset: 25,000+ cybersecurity examples (ReAct format)
- Hardware: NVIDIA A100 80GB / RTX PRO 6000
- Training cost: <$100 USD
Tools Available
- cve_search β Query NVD/CVE databases
- exploit_db β Search Exploit-DB
- web_fetch β Fetch URLs for OSINT
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
"rod123/neos-v10-merged",
torch_dtype="bfloat16",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("rod123/neos-v10-merged")
Intended Use
Security research, CTF competitions, penetration testing assistance, defensive security analysis. Not intended for malicious use.
Links
- Downloads last month
- 13