SAIF v2 Pro — Circuit-Aware Security Specialist
SAIF (سيف) v2 Pro is a unified security analysis model based on Qwen3-8B, trained with Circuit-Aware LoRA and IACR (Integrated Adapter Circuit Refinement).
Architecture
- Base Model: Qwen3-8B (36 layers)
- Circuit-Aware LoRA: Trained on 387 security analysis samples (rank=8, scale=20)
- IACR Fine-tuning: Integrated Adapter Circuit Refinement (loss 0.033)
- 13 Security Probes: Avg accuracy 98.7%
- Fused Format: MLX native (single 15GB model)
Capabilities
| # | Capability | Supported |
|---|---|---|
| 1 | SQL Injection Analysis | ✅ |
| 2 | SSRF Detection | ✅ |
| 3 | XSS Analysis | ✅ |
| 4 | RCE Exploitation | ✅ |
| 5 | IDOR Detection | ✅ |
| 6 | CSRF Analysis | ✅ |
| 7 | LFI Detection | ✅ |
| 8 | Open Redirect | ✅ |
| 9 | CORS Misconfig | ✅ |
| 10 | Auth Bypass | ✅ |
| 11 | Business Logic | ✅ |
| 12 | Race Conditions | ✅ |
| 13 | LLM Prompt Inject | ✅ |
Usage (MLX)
from mlx_lm import load, generate
model, tokenizer = load("BinSaqban/SAIF-v2-Pro")
prompt = "Analyze this SSRF vulnerability: ..."
messages = [{"role": "user", "content": prompt}]
prompt_text = tokenizer.apply_chat_template(messages, tokenize=False)
response = generate(model, tokenizer, prompt=prompt_text, max_tokens=512)
print(response)
Server Mode
mlx_lm.server --model BinSaqban/SAIF-v2-Pro --port 8452
Performance
- Latency: ~7.4s for 200 tokens (27 tok/s on M2 Ultra)
- Accuracy: 98.7% on 13 security probes
- Memory: ~16GB (M2 Ultra 192GB)
Training
Trained on Hayula Labs infrastructure using MLX on M2 Ultra (192GB).
- Circuit-Aware: 500 iterations, train loss 0.927
- IACR: Additional refinement, final loss 0.033
License
Apache 2.0