Apple’s on-device AI toolkit is useful, but not magic
Apple’s official apple/coreai-models repository helps convert some models into Apple’s on-device .aimodel format. It is the companion to Apple’s new Core AI framework, which was shown at WWDC26. The toolkit includes export examples for a selected set of popular models, Python building blocks for making PyTorch models for device use, a Swift package for running those models inside macOS or iOS apps, and that guide a coding assistant on how to use Core AI correctly.
The repository is an official Apple project, uses the BSD-3-Clause license, and had about 1,200 stars while still new. Claims that Apple turned billions of iPhones into local AI machines, or that any model can now run natively on an iPhone, are overstated. This does not mean a huge model like Llama 70B will run easily on a phone.
Apple already had Core ML and coremltools, so this is better understood as a new official path for Core AI rather than Apple’s first move into .
Key points
- apple/coreai-models is an official Apple repository.
- It converts selected models into Apple’s .aimodel format.
- It includes a Swift package for running models in macOS and iOS apps.
- It also includes Python tools for preparing PyTorch models for device use.
- Claims that any model can now run locally on an iPhone are exaggerated.