MLX

MLX is an open-source array framework for machine learning created by Apple machine learning research (the ml-explore organization on GitHub). It is designed for efficient and flexible training and inference of models on Apple silicon. Its Python API closely follows NumPy, with additional interfaces available in C++, C, and Swift.

Key features include unified (shared) memory, where arrays live in shared memory so operations can run on CPU or GPU without transferring data; lazy computation, where arrays are only materialized when needed; dynamic computation graph construction without slow recompilations; and composable function transformations for automatic differentiation, automatic vectorization, and graph optimization. The design is inspired by PyTorch, JAX, and ArrayFire, and the code is released under the MIT license.

MLX is used for tasks such as transformer language model training, text generation (LLaMA, LoRA fine-tuning), image generation (Stable Diffusion), and speech recognition (Whisper). The framework runs locally on-device; beyond Apple silicon, a CUDA backend is also available for Linux.