Core ML Models Zoo

31 PyTorch models converted to Core ML (.mlpackage) for on-device inference on iPhone, iPad and Mac — detection, segmentation, depth, matting, diffusion, TTS, source separation and more. 6.96 GB of packages in one repo.

Every model here has a matching, readable conversion script and — for most of them — a standalone SwiftUI sample app in the CoreML-Models repo.

This repo is also the backing store for the Models Zoo app on the App Store, which downloads and runs any of these models on device without writing code. models.json is that app's manifest — treat it as machine-owned.

Each model also has its own repo with a model card, a demo image and the unzipped .mlpackage — the Model column links there. The In this repo column is the zipped copy the app downloads.

Models

† Downloadable here, but not yet selectable inside the Models Zoo app — the shipped app build has no UI template for them.

Also in the Models Zoo app (hosted in separate repos)

Download

Each model lives in its own directory and ships as a zipped .mlpackage.

hf download mlboydaisuke/coreml-zoo --include "moge2/*" --local-dir ./moge2
unzip './moge2/moge2/*.zip' -d ./moge2

# everything (large)
hf download mlboydaisuke/coreml-zoo --local-dir ./coreml-zoo

Use in Swift

import CoreML

let config = MLModelConfiguration()
config.computeUnits = .cpuAndNeuralEngine   // see each model's card

let compiled = try await MLModel.compileModel(at: mlpackageURL)
let model = try MLModel(contentsOf: compiled, configuration: config)

Two things that bite on real devices, both documented per model in docs/coreml_conversion_notes.md:

1. Compute units are load-bearing. Several models are converted for a specific backend (FP32 + .cpuOnly where FP16 attention overflows, .cpuOnly where the iOS GPU hits the MPS singleton-slice bug). Switching them is not free.
2. MLMultiArray.dataPointer is not contiguous on the Neural Engine. ANE pads rows for SIMD alignment — always read through array.strides.

Licenses

Licenses are per model and inherited from upstream. Several are non-commercial (MatAnyone: S-Lab 1.0, SinSR: CC BY-NC-SA 4.0, RMBG-1.4: Bria RMBG license, Pixelization: research use only) and the YOLO models are AGPL-3.0 / GPL-3.0. Check the table above and the upstream repo before shipping anything commercially.

Credits

Conversions by john-rocky (Daisuke Majima). Upstream authors are credited per model above.

---

More models in this format: Core ML Model Zoo — 46 models, each with the recipe that produced it.

Want a different model on-device? Open a request — free, open weights only; the export and its measured numbers get published publicly.