← Back to Projects

Is Quantization a Deal-breaker? Empirical Insights from Large Code Models

Saima Afrin, Bowen Xu, Antonio Mastropaolo

Prior work on quantizing large code models focuses almost entirely on functional correctness, overlooking how quantization affects code quality. We apply Activation-aware Weight Quantization (AWQ) to CodeLlama and DeepSeek-Coder generating Java and Python code, and use state-of-the-art static-analysis tools to measure quality attributes such as cyclomatic complexity, cognitive complexity, and lines of code. Our findings show that 4-bit quantization not only preserves functional correctness but also retains key qualitative attributes developers care about, such as maintainability and structural simplicity.