kinetigor-dpo-cybersec (Legacy 0.5B)

DEPRECATED -- This is the legacy 0.5B model. For the canonical KIN model, use nyxspecter4/kin-sft-lora (3B, merged weights, GGUF, higher quality). This 0.5B model remains public for backward compatibility and edge deployment on CPU-only devices.

DPO-tuned Qwen2.5-0.5B-Instruct for cyber-defense, penetration testing, and security-analyst reasoning.

Trained on the curated nyxspecter4/kin-cyber-dpo-v2 dataset (1,637 DPO pairs) spanning MITRE ATT&CK, OWASP Top-10 (Web & API 2023), GraphQL attack surfaces, NIST CSF, SANS Top-20, CIS Controls v8, Sigma detection rules, CVE, and CISA KEV.

Ultra-Lightweight Edge Deployment: At only 0.5B parameters (398MB Q4_K_M GGUF), Kinetigor runs at >60 tokens/sec on pure CPU on any MacBook, Raspberry Pi, or phone without requiring a dedicated GPU.

Upgrade to the 3B Canonical

# Ollama - 3B canonical (recommended)
ollama run hf.co/nyxspecter4/kin-sft-lora-gguf:Q4_K_M

# Ollama - 0.5B legacy (edge deployment)
ollama run hf.co/nyxspecter4/kinetigor-dpo-cybersec

Local Inference via Ollama / llama.cpp (Zero GPU Required)

1. Ollama (One-Line Run)

ollama run hf.co/nyxspecter4/kinetigor-dpo-cybersec

2. llama.cpp CLI

llama-cli -hf nyxspecter4/kinetigor-dpo-cybersec -p "Explain how ASLR memory layout randomization prevents buffer overflow exploits:"

3. Python (llama-cpp-python)

from llama_cpp import Llama
from huggingface_hub import hf_hub_download

model_path = hf_hub_download("nyxspecter4/kinetigor-dpo-cybersec", "kinetigor-v6-Q4_K_M.gguf")
llm = Llama(model_path=model_path, n_ctx=2048)

output = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are KIN, an elite cyber defense analyst."},
        {"role": "user", "content": "What is the mandatory structure of a winning bug bounty reproduction step?"}
    ]
)
print(output["choices"][0]["message"]["content"])

Competitive Landscape (Live Probed 2026-09-08)

License & Citations

Licensed under the MIT License.

@misc{kinetigor2026,
  author = {NyxSpecter4},
  title = {Kinetigor: Ultra-Lightweight DPO-Tuned Cyber Defense Model},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/nyxspecter4/kinetigor-dpo-cybersec}}
}