The first available MLX package on PyPI provided a Python entry point for array computation on Apple silicon, with CPU and GPU arrays sharing physical memory.
MLX & MLX LM
Also known as: Apple MLX · MLX LM
Plain English
MLX is an array framework for running and adapting machine learning models on Apple silicon. MLX LM builds on it to run and fine-tune language models on Macs.
Technical Overview
MLX evaluates arrays lazily and lets CPU and GPU operations access the same underlying memory on Apple silicon. MLX LM uses this model for local generation and supports weight Quantization and adapter fine-tuning. MLX also exposes distributed collectives, but the interconnect and choice of sharding strategy affect performance across machines.
Timeline (4)
2023
Release2024
ReleaseMLX 0.10 added asynchronous evaluation and faster quantized matrix-vector operations, improving the execution path for local language model generation.
2025
EcosystemApple demonstrated native Swift generation with a quantized Mistral model and a reusable key-value cache, showing how a macOS application can generate across multiple turns in one process.
2026
EcosystemApple demonstrated MLX LM sharding models across Macs using tensor and pipeline parallelism and RDMA over Thunderbolt 5. Sharding allows model weights to span machines, while RDMA reduces communication overhead between them.
See also (9)
Open full graphMLX & MLX LM