Swift-AI is a machine learning library written entirely in Swift, bringing neural networks natively to iOS and macOS. It was one of the earliest serious attempts to do deep learning in Apple’s language, years before the current generation of on-device ML tooling. The project has about 6,000 stars on GitHub and is released under the MIT license.
The library implements feed-forward neural networks with backpropagation training directly in Swift, with a modular design that separates the core network code from supporting utilities. Because it is pure Swift with no Python or server component, models can be trained and run inside an app itself — an appealing idea in 2016–2017, when the repository was most active (its last commit dates to 2017). Today Swift-AI is best understood as a compact, readable reference implementation: small enough to study end to end, and a piece of Swift ML history from before Core ML became the standard route for on-device models.
The project is also a small time capsule of early mobile ML: it arrived when doing deep learning on a phone still meant writing the math yourself, and its straightforward Swift code — matrix operations, activation functions, gradient descent — shows exactly what a neural network library contains under the hood. For teaching and experimentation, that transparency is the point. Anyone building a production Apple app today would normally reach for Core ML or Apple’s newer on-device frameworks, but as a learning resource Swift-AI remains unusually approachable.
Free. Swift-AI is open source under the MIT license.
Swift developers curious about the fundamentals of neural networks, and anyone maintaining or studying early on-device ML code for Apple platforms.