LUT-KAN: An open software toolkit for LUT-compiled and quantized CPU inference of Kolmogorov–Arnold networks

Kolmogorov–Arnold Networks (KANs) offer interpretable alternatives to multilayer perceptrons through learnable univariate activation functions on edges. However, their reliance on B-spline evaluation creates computational bottlenecks that limit practical deployment. We present LUT-KAN, an open-source Python toolkit that compiles trained KAN models into lookup table (LUT) representations for efficient CPU inference. The toolkit implements segment-wise quantization with configurable precision, linear interpolation for output reconstruction, and provides both NumPy and Numba computational backends. Our controlled experiments on synthetic benchmarks demonstrate speedups of 8–14× over equivalent B-spline baselines while maintaining approximation errors below 10−3 with 64 samples per segment. We validate practical applicability through a denial-of-service detection case study using a multi-layer KAN architecture, achieving 0.99 F1-score with minimal accuracy degradation. LUT-KAN is freely available under MIT license with comprehensive documentation, reproducible benchmarks, and configuration-driven experiment management.