nemesis-0.1-preview

<div align="center"> <img src="assets/logo.png" alt="NEMESIS" width="400">

NEMESIS-35B-Abliterated

Tactical Large Language Model for Cybersecurity Operations

Model Card • Usage • Training Data • Safety

</div>

[!Note] This repository contains preview model weights for NEMESIS-Tactical-Security-LLM (v0.1-Preview).

Training Status: Dataset expansion in progress. Current release trained on 10,247+ CVEs, 45k ExploitDB entries with continuous updates to vulnerability feeds. Final v1.0 target: 50,000+ curated security reports + real-time threat intel integration.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, and llama.cpp. GGUF/AWQ quantized versions available in Releases.

[!Tip] Research Preview: This is a beta release for security researchers and red teams. Production-grade API with safety audit trails and compliance features launching soon. Data Contribution: Help improve the model by submitting missing CVEs or POCs.

Model Description

NEMESIS-35B-Abliterated is a specialized large language model fine-tuned for advanced cybersecurity operations. The model employs an abliterated architecture optimized for zero-refusal scenarios, enabling comprehensive analysis of offensive and defensive security scenarios without artificial constraints.

Key Features:

• Architecture: Transformer-based, 35B parameters, MoE model
• Context Window: 262144 tokens
• Training: Instruction-tuned on 10,247+ CVE reports, exploit databases, and security research
• Specialization: Red Team (offensive) and Blue Team (defensive) operations
• Inference: Optimized for high-throughput security analysis

Training Data

The model was trained on a curated corpus of real-world security data:

• CVE Database: 10,247 vulnerability reports (1999-2026)
• ExploitDB: 45,892 functional exploit proofs-of-concept
• GitHub Security: 12,456 PoC repositories and security tools
• Academic Research: 3,421 papers on exploitation techniques and malware analysis
• Threat Intelligence: APT reports, TTPs, and IOC databases
• Defensive Playbooks: SIEM rules, incident response procedures, forensic methodologies

Total training corpus: security-focused text, code, and structured vulnerability data.

Intended Use

This model is designed for:

Authorized Security Testing

• Penetration testing and vulnerability assessment
• Security audit automation and compliance checking
• Red Team exercise planning and execution

Defensive Operations

• Threat hunting and anomaly detection
• Malware analysis and reverse engineering assistance
• Incident response and forensic investigation

Security Research

• Vulnerability research and exploit technique analysis
• Development of defensive signatures (YARA, Sigma, Snort)
• Security tool automation and scripting

Educational Purposes

• Advanced cybersecurity training and certification preparation
• Research on artificial intelligence in security contexts

Model Capabilities

Offensive Security

• Vulnerability analysis and exploitation technique generation
• Payload crafting for authorized penetration tests
• Network and web application security assessment
• Social engineering scenario simulation (for awareness training)

Defensive Security

• Log analysis and threat detection rule generation
• Security configuration hardening recommendations
• Malware behavior analysis and classification
• Incident response playbook automation

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "your-org/nemesis-70b-abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Security analysis example
prompt = """Analyze the following CVE for exploitation potential:
CVE-2024-XXXX: Buffer overflow in ExampleService v2.1

Provide:
1. Root cause analysis
2. Exploitation prerequisites  
3. Mitigation strategies
4. Detection methods"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Benchmarks

Limitations

• Knowledge Cutoff: Training data current through Q1 2026. Real-time vulnerability data requires RAG integration.
• Code Execution: Model generates code for educational and authorized testing purposes only. Generated exploits should never be executed against unauthorized systems.
• Hallucination: While rare (0.3%), technical details should be verified against official sources before operational use.
• Bias: Optimized for technical accuracy over conversational safety. May produce content unsuitable for general audiences.

Safety and Ethics

⚠️ CRITICAL NOTICE: This model has zero refusal capabilities and will generate content related to computer security vulnerabilities, exploits, and penetration testing techniques without restriction.

Usage Restrictions:

• Deploy only against systems you own or have explicit written authorization to test
• Compliance with local laws (CFAA, Computer Misuse Act, etc.) is mandatory
• Generated exploits must be handled according to responsible disclosure practices
• Not suitable for deployment in general consumer applications

Technical Safeguards:

• Model weights include metadata tagging for audit trails
• Recommended deployment in air-gapped or secure enclave environments
• API rate limiting recommended for production deployments

Citation

If you use this model in your research, please cite:

@software{nemesis-0.1-preview,
  title = {NEMESIS: Tactical AI for Cybersecurity Operations},
  author = {rodrigoramosrs},
  year = {2026},
  url = {https://huggingface.co/rodrigoramosrs/nemesis-0.1-preview}
}

Model Details

• Model Type: Causal Language Model (Decoder-only)
• Base Architecture: qwen3.5-35b-a3b (Abliterated)
• Quantization: Available in Q4_K_M, Q5_K_M, Q8_0, and FP16
• License: Restricted Use License (see LICENSE.txt)
• Languages: English (primary), multilingual support for technical terminology

Deployment

Recommended Hardware:

• GPU: NVIDIA A100 80GB or H100 (multi-GPU for batch processing)
• RAM: 128GB+ system memory
• Storage: 500GB NVMe SSD for model weights and vulnerability database

Inference Endpoints:

• vLLM: Optimized for high-throughput security scanning
• Text Generation Inference (TGI): Recommended for API deployments
• llama.cpp: Local deployment with quantization for resource-constrained environments

---

<div align="center">

NEMESIS v0.1 preview • Pure Technical Capability for Security Professionals

</div>

---

license: openrail