Apple’s Core AI toolkit ships model recipes and coding-agent plugins

What it does This is Apple’s official companion to its Core AI framework. It provides export recipes that convert popular open-source models—largely from Hugging Face—into .aimodel bundles for on-device inference on macOS and iOS. A Swift package supplies runtime utilities for app integration, while Python primitives help author custom PyTorch models.

The interesting bit The repo also distributes agent plugins for Claude Code, OpenAI Codex CLI, and Gemini CLI that teach agents to export models, follow PyTorch authoring rules, and explore compression configs. Apple is keeping contribution strictly read-only at launch: pull requests will be closed, though bug reports and model requests are welcome.

Key highlights

• Export recipes for popular open-source models to Core AI .aimodel format
• Swift runtime package (coreai-models) for integrating models into macOS/iOS apps via the Core AI framework
• Python primitives and utilities for authoring custom PyTorch models and exporting them
• Agent skills for Claude Code, Codex CLI, and Gemini CLI to automate Core AI workflows
• Curated gallery approach: not accepting code contributions at launch; issues and model requests only

Caveats

• Pull requests are explicitly closed at launch; only issues and model requests are accepted
• Some models need extra resources (e.g., tokenizers, multi-model pipelines) beyond the standalone .aimodel file
• Requires macOS and iOS 27.0+ and Xcode 27.0+ per the README

Verdict Worth bookmarking if you are building on-device AI for Apple platforms and want official export recipes and Swift glue. Skip it if you are itching to contribute code immediately or need cross-platform support.