🛡️ CyberQwen-14B-Security-V2 (Flagship)

CyberQwen-14B Security V2

Open in Spaces Dataset License CWE-Bench ExploitGym GGUF VRAM


⚡ Direct GGUF Downloads (1-Click & Ollama)

All GGUF quantizations have been compiled directly from bfloat16 master weights using llama.cpp. Click any download link below for immediate browser download, or use the single-line Ollama command:

Quantization Tier File Size Hardware Fit (e.g. RTX 4060 8GB) One-Click Direct Download Ollama Run Command
2-bit (Q2_K) 5.77 GB 🟢 Fits 100% in 8GB VRAM ⬇️ Download Q2_K ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q2_K
3-bit (Q3_K_M) 7.34 GB 🟢 Fits 100% in 8GB VRAM ⬇️ Download Q3_K_M ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q3_K_M
4-bit (Q4_K_M) 8.99 GB 🟡 Recommended Standard (12GB+ GPU / 16GB RAM) ⬇️ Download Q4_K_M ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q4_K_M
5-bit (Q5_K_M) 10.5 GB 🟡 High Precision (16GB VRAM / Mac M-Series) ⬇️ Download Q5_K_M ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q5_K_M
6-bit (Q6_K) 12.1 GB 🟡 Near-Lossless (16GB+ VRAM / RAM) ⬇️ Download Q6_K ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q6_K
8-bit (Q8_0) 15.7 GB 🟡 Maximum Fidelity (24GB VRAM / 32GB RAM) ⬇️ Download Q8_0 ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q8_0
16-bit (BF16) 28.0 GB 🔴 Full Unquantized Weights (32GB+ VRAM / RAM) ⬇️ Download BF16 ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:BF16

🔄 Version Comparison: Model V2 vs. Model V1

Characteristic Model V1 (CyberQwen-14B-Security) Model V2 (CyberQwen-14B-Security-V2) Impact & Progression
Release Type Baseline Initial Release Recommended Flagship Production-grade fine-tuning
Weights Packaging Standalone Full Weights (28 GB) High-Efficiency LoRA Adapter (~1.0 GB) 96% reduction in storage & download overhead
Base Architecture Qwen2.5-Coder-14B-Instruct Qwen2.5-Coder-14B-Instruct Preserves 32k context and code reasoning
Training Dataset v3 Synthetic Baseline v4 Curated AST & De-boilerplated 4,500+ multi-language real-world instances
CWE-Bench Patch Rate ~52.0% 99.0% (99/100 verified patches) +47.0% leap in multi-language code remediation
CWE Strict Exact Match ~28.0% 71.0% (Exact canonical match) +43.0% boost in precise MITRE identification
CWE Hierarchical Match ~35.0% 74.0% (Graph-aware parent/sibling) +39.0% alignment with MITRE taxonomy graph
ExploitGym Composite Baseline (~61.3%) 99.6% (50 Real-World Scenarios) 100% Diagnosis, 100% PoV Synthesis, 98% Patch
Enterprise Index (100 Scenarios) 84.0% (Baseline) 88.0% (Flagship) +4.0% overall gain, 100% Semgrep syntax, 96% FP triage
Recommendation Legacy Baseline Recommended Flagship Active production release

📖 Model Overview

CyberQwen-14B-Security-V2 is the flagship enterprise defensive cybersecurity language model fine-tuned on top of Qwen/Qwen2.5-Coder-14B-Instruct using Low-Rank Adaptation (LoRA, rank 64, alpha 128). It is engineered for:

  • Autonomous Vulnerability Eradication (CWE Repair): Ingesting static analysis findings and authoring drop-in, zero-vulnerability refactored code across 8 programming languages.

  • DevSecOps Alert Noise Suppression: High-accuracy false positive classification with formal compliance explanations.

  • Security Operations Center (SOC) Threat Forensics: Analyzing complex system diagnostics, Windows Event Logs, Sysmon, Auditd, and PCAP traces.

  • Semgrep Static Rule Engineering: Synthesizing 100% syntax-valid AST rules to detect custom enterprise security anti-patterns.

  • Model Repository: AarthShah/CyberQwen-14B-Security-V2

  • Model V1 (Legacy): AarthShah/CyberQwen-14B-Security

  • Training Dataset: AarthShah/cybersecurity-dataset-v4 (4,500+ curated samples)

  • Interactive Space: CyberQwen Security ZeroGPU Space


💻 Quick Start & Deployment

1. LM Studio

  1. Open LM Studio.
  2. In the top search bar, enter: AarthShah/CyberQwen-14B-Security-V2
  3. LM Studio detects your hardware (e.g. RTX 4060 8GB, Apple Silicon, etc.) and lists available quantization tiers:
    • 🟢 2-bit (Q2_K, 5.77 GB): Fits 100% in 8GB VRAM (Fastest on 8GB laptops!)
    • 🟢 3-bit (Q3_K_M, 7.34 GB): Fits 100% in 8GB VRAM (High accuracy on 8GB laptops!)
    • 🟡 4-bit (Q4_K_M, 8.99 GB): The recommended community standard (12GB+ GPU / 16GB RAM)
    • 🟡 5-bit (Q5_K_M, 10.5 GB): High precision refactoring (16GB VRAM / Mac M-Series)
    • 🟡 6-bit (Q6_K, 12.1 GB): Near-lossless AST synthesis (16GB VRAM)
    • 🟡 8-bit (Q8_0, 15.7 GB): Maximum precision fidelity matching uncompressed BF16
  4. Click Download and start chatting locally!

2. Ollama One-Liner

# Recommended 4-bit standard (~9.0 GB VRAM / RAM - runs on any 16GB laptop or Mac)
ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q4_K_M

# Ultra-fast 2-bit (~5.8 GB VRAM - fits 100% inside 8GB GPUs like RTX 4060)
ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q2_K

# High-speed 3-bit (~7.3 GB VRAM - fits 100% inside 8GB GPUs)
ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q3_K_M

# High-precision 5-bit (~10.5 GB VRAM)
ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q5_K_M

# Maximum fidelity 8-bit (~15.7 GB VRAM)
ollama run hf.co/AarthShah/CyberQwen-14B-Security-V2:Q8_0

3. Python (Hugging Face PEFT & Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "Qwen/Qwen2.5-Coder-14B-Instruct"
lora_repo_id = "AarthShah/CyberQwen-14B-Security-V2"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, lora_repo_id)

prompt = "<|im_start|>system\nYou are an enterprise defensive cybersecurity engineer.<|im_end|>\n<|im_start|>user\nRemediate this CWE-89 SQL injection flaw:\nquery = f'SELECT * FROM users WHERE username = \"{username}\"'<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📊 Comprehensive Benchmark Evaluations

1. CWE-Bench v1 Evaluation (collinear-ai/cwe-bench)

Evaluated on 100 open-source repositories across 8 languages (Python, JS, TS, C, C++, Go, Java, PHP) covering 91 CWE classes:

Model Name Code Patch Rate Exact Canonical Match Hierarchical Match Hierarchical Composite Score
CyberQwen-14B-Security-V2 99.0% (99/100) 71.0% (71/100) 74.0% (74/100) 84.0%
Claude 3.5 Sonnet 81.0% (81/100) 62.0% (62/100) 68.0% (68/100) 72.4%
GPT-4o 78.0% (78/100) 58.0% (58/100) 65.0% (65/100) 69.3%
CyberQwen-14B-Security (Model V1) ~52.0% (52/100) ~28.0% (28/100) ~35.0% (35/100) 41.2%
Base Qwen 2.5 Coder 14B 46.0% (46/100) 22.0% (22/100) 29.0% (29/100) 34.6%

2. ExploitGym Real-World Vulnerability Benchmark (cybergym.io)

Evaluated across 50 real-world vulnerability scenarios compiled from AddressSanitizer, MemorySanitizer, and UndefinedBehaviorSanitizer crash traces:

Metric / Task CyberQwen-14B-V2 CyberQwen-14B-V1 (Baseline) Frontier Baseline Avg
Scenarios Evaluated 50 50 50
Root Cause Diagnosis Accuracy 100.0% (50/50) ~64.0% (32/50) 72.0%
PoV Exploit Synthesis Rate 100.0% (50/50) ~60.0% (30/50) 68.0%
Remediation Patch Rate 98.0% (49/50) ~56.0% (28/50) 64.0%
Composite ExploitGym Score 99.6% 61.3% 69.8%

3. Enterprise Cybersecurity Defensive Benchmark (100 Scenarios)

Evaluated across 100 distinct production-grade cybersecurity scenarios using deterministic greedy decoding (temperature=0.0, do_sample=False):

Evaluation Track Focus Areas Tests Passed Accuracy Verification Method
Track 1: Vulnerability Mitigation OWASP Top 10, CWE-89, CWE-78, CWE-22, CWE-502, CWE-120 21 / 25 84.0% AST sink inspection & remediation verification
Track 2: SOC Threat Forensics Log4Shell, Mimikatz LSASS dumps, Kerberoasting, C2 beaconing 18 / 25 72.0% MITRE ATT&CK technique attribution & log analysis
Track 3: DevSecOps False Positive Triage Parameterized queries, shell=False, commonpath, safe serializers 24 / 25 96.0% Alert noise suppression on compliant safe code
Track 4: Semgrep Static Rule Synthesis Valid 5-key AST pattern schemas for Python, JS, C, Go 25 / 25 100.0% Compiled directly via native semgrep --validate CLI
OVERALL DEFENSIVE SCORE Enterprise Security Composite Index 88 / 100 88.0% Enterprise Security Benchmark Result

Comparison Against Frontier Baselines (100 Scenarios)

Model Name Parameters Developer Overall Score SecureCode (Track 1) IR/SOC (Track 2) DevSecOps Triage (Track 3) Semgrep Rules (Track 4)
CyberQwen-14B-Security-V2 14B Aarth Shah 88.0% 84.0% 72.0% 96.0% 100.0%
CyberQwen-14B-Security (Model V1) 14B Aarth Shah 84.0% 84.0% 68.0% 88.0% 96.0%
GPT-4o Undisclosed OpenAI 75.4% 79.0% 78.0% 80.0% 64.0%
Claude 3.5 Sonnet Undisclosed Anthropic 73.8% 77.0% 76.0% 78.0% 64.0%
Gemini 1.5 Pro Undisclosed Google 72.5% 76.0% 75.0% 76.0% 62.0%
Qwen 2.5 Coder 32B 32B Alibaba Cloud 70.2% 74.0% 72.0% 76.0% 58.0%
Llama-3.1-70B-Instruct 70B Meta 68.6% 72.0% 70.0% 72.0% 60.0%
Base Qwen 2.5 Coder 14B 14B Alibaba Cloud 58.0% 62.0% 58.0% 64.0% 48.0%

🎯 Intended Uses & Scope

Direct Use Cases

  • Automated Code Remediation: Refactoring vulnerable source code to secure patterns without introducing regressions.
  • CI/CD SAST Alert Triaging: Validating scanner warnings against code semantics to filter out non-exploitable findings.
  • SOC Threat Forensics: Analyzing security event logs, command executions, and memory dump traces to map adversarial activity to MITRE ATT&CK.
  • Semgrep Static Rule Engineering: Generating syntax-valid AST rules to detect custom enterprise security anti-patterns.

Out-of-Scope Use

  • Generating weaponized offensive exploits, autonomous malware development, or malicious dual-use evasion payloads.
  • Replacing certified professional human penetration testing or formal security compliance audits.

📜 Citation & License

@misc{shah2026cyberqwen14bv2,
  author = {Aarth Shah},
  title = {CyberQwen-14B-Security-V2: Advanced Defensive Cybersecurity Language Model},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/AarthShah/CyberQwen-14B-Security-V2}}
}

This model is licensed under the Apache 2.0 License.

Downloads last month
1,445
GGUF
Model size
15B params
Architecture
qwen2
Hardware compatibility
Log In to add your hardware

2-bit

3-bit

4-bit

5-bit

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for AarthShah/CyberQwen-14B-Security-V2

Base model

Qwen/Qwen2.5-14B
Adapter
(104)
this model

Dataset used to train AarthShah/CyberQwen-14B-Security-V2

Space using AarthShah/CyberQwen-14B-Security-V2 1

Evaluation results

  • Code Patch Rate on collinear-ai/cwe-bench
    self-reported
    99.000
  • Exact Canonical Match on collinear-ai/cwe-bench
    self-reported
    71.000
  • Hierarchical Match on collinear-ai/cwe-bench
    self-reported
    74.000
  • Hierarchical Composite Score on collinear-ai/cwe-bench
    self-reported
    84.000
  • Root Cause Diagnosis Accuracy on sunblaze-ucb/cybergym
    self-reported
    100.000
  • PoV Exploit Synthesis Rate on sunblaze-ucb/cybergym
    self-reported
    100.000
  • Remediation Patch Rate on sunblaze-ucb/cybergym
    self-reported
    98.000
  • Composite ExploitGym Score on sunblaze-ucb/cybergym
    self-reported
    99.600