This research evaluates the safety implications of LLM quantization techniques that enable deployment on resource-constrained devices.

• Different quantization methods significantly affect model safety performance
• The newly created OpenSafetyMini dataset benchmarks safety across quantized models
• 2-bit quantization drastically increases harmful outputs (up to 66% in some cases)
• Models with 4-bit precision maintain reasonable safety while reducing resource requirements

For security professionals, this research provides crucial guidance on quantization trade-offs, demonstrating that aggressive compression can compromise safety guardrails while identifying more balanced approaches for secure deployment.

Investigating the Impact of Quantization Methods on the Safety and Reliability of Large Language Models