Apple has released MLX as an Apple silicon optimised Array framework.
Users have observed up to 2x faster inference times for examples problems such as MNIST (Twitter)