KANX: A Production-Grade Open-Source Library for Kolmogorov-Arnold Networks

Kolmogorov–Arnold Networks (KANs) are a recently proposed neural architecture in which every weight
is replaced by a learnable univariate B-spline, grounded in the Kolmogorov–Arnold representation
theorem. Despite growing academic interest, the ecosystem has lacked a production-ready, multi-backend
implementation suitable for both rigorous experimentation and real-world deployment. We present KANX
(kanx), the first fully-documented, cross-framework KAN toolkit: a pip-installable library (TensorFlow
primary, PyTorch secondary) with real ONNX export, a FastAPI REST service, Docker/Kubernetes
manifests, a 113-test suite (94% coverage), continuous integration, and automated PyPI releases. Using
KANX's reproducible benchmark harness, we conduct a rigorous, multi-baseline comparison of KANs
against parameter-matched MLPs. On a smooth 2-D synthetic regression target—the best-case regime for
KAN theory—a KAN [2,16,1] achieves a test MSE of 2.14 × 10⁻⁵ with only 432 parameters. We further
report five-fold cross-validated results on the UCI Diabetes dataset, with KAN-TF achieving R² = 0.449 ±
0.130 with only 1,068 parameters—competitive with Ridge regression and substantially ahead of a
parameter-matched MLP (R² = 0.089). All benchmarks were validated on a T4 GPU (Google Colab, June
2026). We explicitly characterize the regime boundaries: KANs excel on smooth, low-dimensional,
separable targets but incur 3–5× higher inference latency than equivalent MLPs. KANX closes the tooling
gap, enabling fair comparisons and broader adoption of KAN research