Apple has opened a repository that treats coding agents as first-class users of its on-device AI stack, shipping agent skills right next to the model export recipes and Swift runtime code. That detail sits quietly in a table near the bottom of the README for coreai-models, but it says a lot about who Apple expects to be exporting models to Core AI format.
What the repo actually contains
The apple/coreai-models repo is organized around four components, each mapped to a directory. models/ holds a model catalog with export recipes for pulling popular open source models from Hugging Face and other sources into Core AI format. python/ provides PyTorch primitives for authoring custom Core AI models plus utilities for the export step. swift/ is a Swift package, also named coreai-models, that gives you runtime utilities to run those models inside a macOS or iOS app. skills/ holds plugins that let coding agents work with Core AI.
The export step produces standalone .aimodel files. The README is candid that a single file is often not the whole story: language models need a tokenizer, and diffusion models run several models in sequence as one pipeline. For those cases the recipes emit a resource folder containing one or more .aimodel files alongside whatever else is required, and the Swift package handles wiring that into an app. There are also CLI tools for running exported models directly on a Mac, with per-model READMEs documenting the available invocations.
The requirements are strict and current
This is not a "works on anything" project. Running and integrating models requires macOS and iOS 27.0 or later and Xcode 27.0 or later. The Python side leans on uv; the README's first instruction is to install it through Homebrew or the Astral install script, then follow the per-model README for each export recipe. To see what is supported, you clone the repo and run uv run coreai.model.registry --list-models. The narrow version floor tells you this is aimed at the newest Apple platforms rather than a broad back-catalog of devices.
Skills aimed at coding agents
The part I find most telling is the skills/ plugin. Apple ships three named skills. working-with-coreai covers the end-to-end path of exporting a PyTorch model and running it on the Core AI runtime. model-authoring is described as a set of empirical rules for writing PyTorch models that run on-device, touching BC1S layout, op compatibility, KV cache patterns, precision rules, and mixture-of-experts. model-compression-exploration walks through quantization and palettization configurations using a tool the README calls coreai-opt.
The install instructions cover Claude Code, Codex CLI, and Gemini CLI, each registering the repo as a plugin marketplace or extension. Publishing the same skills across three different agent front ends is a deliberate bet that people will lean on an agent to handle the fiddly parts of on-device model work, and that Apple would rather encode its own "empirical rules" than let agents guess.
A curated gallery, not an open project
The contribution policy is unusually direct. Apple states plainly that it is not accepting code contributions at launch, and that any pull request will be closed. The stated reason is scope: the maintainers want a curated, well-tested gallery of models and a reliable Swift package while they learn how the community uses the project. They do welcome GitHub Issues for bug reports and model requests. Read that alongside the repo's youth, created in June 2026, and the picture is a controlled first release rather than a community free-for-all.
For anyone building on-device AI in the Apple ecosystem, coreai-models is worth watching precisely because it bundles the export recipes, the runtime code, and the agent tooling in one BSD-3-Clause repository. The claims about supported models and compression are the authors' own, and the model catalog is only as good as what lands in models/ over time. But the structure signals intent: Apple wants the export-to-device path to be something an agent can drive, on current hardware, with Apple's rules baked in.
GitHub: https://github.com/apple/coreai-models
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)