Creator of Apple's MLX Framework Departs
Awni Hannun, the creator of Apple's MLX machine learning framework for Apple Silicon, has departed the company. In his announcement, he praised the AI potential of Apple's custom chips and confirmed he is handing off the project to the existing team, signaling confidence in its future.
MLX was first released as an open-source project in December 2023, positioned as a machine learning framework from Apple's research division specifically for developers and researchers. Its design is heavily inspired by frameworks like NumPy, PyTorch, and JAX, offering familiar APIs in Python, C++, Swift, and C to ease adoption. The framework's core advantage lies in its optimization for the unified memory architecture of Apple Silicon. This allows both the CPU and GPU to operate on shared memory without the need for data transfers, a key architectural difference from traditional discrete GPU setups. Key technical features of MLX include lazy computation, where computation graphs are built dynamically and only executed when a result is explicitly requested. This, combined with composable function transformations for automatic differentiation and vectorization, is designed for efficiency in model training and research. Before his tenure at Apple, Awni Hannun was a research scientist at Facebook AI Research and Baidu's Silicon Valley AI Lab, where he co-led the Deep Speech projects. His background also includes a Ph.D. from Stanford University and research focused on speech recognition and privacy-preserving machine learning. MLX fits into Apple's broader strategy of enabling powerful on-device AI, complementing the existing Core ML framework used for model deployment and inference. While Core ML is optimized for running models efficiently within apps, MLX provides the toolchain for the research and training phases directly on Apple hardware. The framework has demonstrated its capability by supporting the training of transformer language models, fine-tuning with techniques like LoRA, and running models such as Stable Diffusion for image generation and OpenAI's Whisper for speech recognition. A community has also begun converting popular models to the MLX format on platforms like Hugging Face.
