@Alexis Manin Yes, it could be. KMath in general is about API, not implementation. There are KMath-core multiplatform implementations that do not guarantee the best performance, but work without additional libraries, but there are also pluggable implementation bridges. MultiK could be (and probably will be) one of such implementations alongside ND4J and Viktor. Currently
@Aleksei Dievskii's measurements show that KMath-core is still slightly faster performance-wise, but in future, it will make sense to add bindings. See
https://github.com/mipt-npm/kmath/issues/186