The latest release of ml-explore/mlx is v0.32.4. MLX is a machine learning array framework for Apple silicon (Apple's own in-house chip), designed by Apple machine learning research so that researchers can train and deploy models efficiently. The source for this release is the repository's GitHub release page.
MLX is a model inference-based framework for handling array and neural network operations on Apple silicon. It takes advantage of both Apple's unified memory and CPU·GPU execution. It offers a NumPy-like Python API and C++, C, and Swift APIs that follow it. It supports automatic differentiation (a feature that computes derivative values in code), automatic vectorization, and computation graph optimization, and covers lazy computation (deferring operations until a result is needed), dynamic graphs, CPU·GPU operations, and unified memory.
The examples are collected in a separate MLX examples repository. There you can find Transformer language model, LLaMA, LoRA, Stable Diffusion, and OpenAI's Whisper examples. The license is MIT, the primary language is C++, and the repository records 28866 stars.
However, what v0.32.4 changed is not documented in the given source records. Neither the change items nor the release date are presented, so the only thing verifiable here is the version number. No performance measurements or execution results are given either, and the score record contains only Momentum 4.6. How the CPU·GPU operation and unified memory paths described in the README differ in this version must be checked directly against the release notes.