AXLearn is Apple’s open-source deep learning library, built on JAX, for training large-scale models with a strongly modular, object-oriented design. Released under the Apache-2.0 license with more than 2,300 GitHub stars and active development in 2026, it is the framework Apple uses and shares for research and production model training across accelerators.
Most deep learning frameworks grow by accretion: layers, trainers and data pipelines tangle together. AXLearn enforces a strict config-driven composition model in which every component โ layers, optimizers, trainers, data inputs โ is a configurable module with an explicit interface. That discipline pays off at scale: the same code trains models from small experiments to very large language and vision models across TPU and GPU clusters, with support for techniques like sequence parallelism and rematerialization. It ships with implementations of transformer architectures, state-space models and the surrounding training infrastructure.
AXLearn is free and open source under the Apache-2.0 license, with no commercial edition.
AXLearn targets ML researchers and platform engineers training large models on JAX who want a codebase that stays maintainable as experiments grow into production training runs.
AXLearn is not just a research artifact: Apple has publicly described using it to train its foundation models on large TPU clusters, and the library reflects that operational heritage with robust checkpointing, data-parallel and model-parallel training strategies, and careful support for numerics at scale. Its configuration system makes experiments reproducible by construction โ every run is fully described by composable config objects โ which appeals to teams that need to audit exactly how a model was produced.