Researcher eyes Qwen 3.6 for offensive security model

Fine-tuning the Qwen 3.6-35B-A3B model for offensive security represents a strategic shift toward high-performance, uncensored local agents for red-teaming. The model is an optimal base due to its hybrid Gated DeltaNet architecture and superior Terminal-Bench scores, which are essential for CLI-heavy security tasks. Researchers can utilize 4-bit QLoRA on RTX 6000 Ada hardware and synthetic data generation from larger teacher models to convert raw logs into high-quality instruction pairs. This approach bypasses the ethical refusal filters common in hosted APIs while emphasizing dataset quality over raw volume to achieve better agentic results.