Ash Vardanian has released NumKong, a comprehensive collection of over 2,000 mixed precision SIMD kernels spanning seven programming languages. The project, formerly known as SimSIMD, represents three years of development with 900 commits and 200,000 lines of code. Comparable in scale to OpenBLAS, NumKong supports everything from Float6 to Int4 precision, enabling high-performance numerical computations across diverse hardware architectures.

NumKong delivers impressive performance benchmarks, outperforming NumPy and PyTorch in several precision ranges. The library includes specialized kernels for geospatial calculations, mesh alignment, and late interaction scoring, offering up to 5,300x speed improvements over existing solutions. It supports cutting-edge instruction sets including RISC-V Vector Extensions, Intel AMX, and Arm SME Tiles, making it adaptable to modern computing hardware.

Currently powering vector search in major databases like ClickHouse, DuckDB, and ScyllaDB, NumKong runs on potentially billions of devices worldwide. The library's development reflects the growing need for specialized numerical kernels that balance speed with accuracy across different precision formats.