On-device models
Models you can run on iPhone, iPad and Mac, from Hugging Face
Core ML and MLX models published on Hugging Face, plus the models Apple publishes there, sorted by downloads in the last 30 days. Core ML models run on iPhone, iPad and Mac through Apple's Core ML framework; MLX models are made for Apple's MLX library, mostly on Apple silicon Macs (MLX Swift also runs on iPhone and iPad). Each name links to the model's page on Hugging Face.
1,009 Core ML models, 1,693 MLX models and 143 from Apple, updated daily from Hugging Face's public API. Last updated 2026-11-02 (UTC).
New this month
The most liked models created in the last 30 days, across Core ML, MLX and Apple's own.
- ConwayResearch/Underdog-Bark-3.0created 2026-10-15 · 86 likes
- FermionResearch/Phonon-2.5created 2026-10-22 · 46 likes
- ConwayResearch/Underdog-Bark-3.0-Procreated 2026-10-15 · 44 likes
- nota-ai/Nemotron-3.5-Lightning-30B-A3B-Nota-Ternary-MLX-1014created 2026-10-12 · 41 likes
- FluidInference/embeddinggemma-2-coremlcreated 2026-10-09 · 32 likes
- RecursiveSI/flint-int4created 2026-10-21 · 28 likes
Showing 13 of 13.
| Model | Task | Size | Downloads (30 days) |
|---|---|---|---|
| apple/coreml-sam2.1-tinyCore ML | Mask generation | — | 2,393 |
| apple/coreml-sam2.1-smallCore ML | Mask generation | — | 2,027 |
| apple/coreml-sam2.1-baseplusCore ML | Mask generation | — | 1,054 |
| apple/coreml-sam2.1-largeCore ML | Mask generation | — | 414 |
| AhmadYarAI/EdgeTAM-CoreML-VideoCore ML | Mask generation | — | 267 |
| mlboydaisuke/MobileSAM-CoreMLCore ML | Mask generation | — | 212 |
| apple/coreml-sam2-tinyCore ML | Mask generation | — | 201 |
| AllanVester/SAM3.1-CoreML-FP16Core MLfrom facebook/sam3.1 | Mask generation | — | 60 |
| apple/coreml-sam2-smallCore ML | Mask generation | — | 46 |
| gumka/sam2.1-small-coremlCore MLfrom facebook/sam2.1-hiera-small | Mask generation | — | 34 |
| apple/coreml-sam2-largeCore ML | Mask generation | — | 30 |
| heypatkay/solume-sam3-coremlCore MLfrom facebook/sam3 | Mask generation | — | 19 |
| apple/coreml-sam2-baseplusCore ML | Mask generation | — | 15 |
Browse by task
Pages for tasks with at least 5 models, across Core ML, MLX and Apple's own.
- Text generation842
- Vision language436
- Speech recognition253
- Text to speech146
- Image generation102
- Text classification50
- Embedding48
- Image to image48
- Image classification33
- Voice activity detection29
- Audio classification27
- Multimodal24
- Sentence similarity24
- Image captioning22
- Named entity recognition22
- Object detection21
- Zero-shot image classification21
- Depth estimation20
- Image embedding18
- Image segmentation17
- Audio enhancement16
- Zero-shot text classification14
- Mask generation13
- Reranking11
- Translation11
- Image to 3D8
- Text to audio7
- Video generation7
- Fill-mask5
More on on-device models
How to run these
- Core ML: add the .mlpackage to your Xcode project, or download a compiled model when the app first needs it, and load it with the Core ML framework. Some repos ship their own Swift package; the model card says which.
- MLX: on a Mac with Apple silicon, the mlx-lm and mlx-audio Python packages load these by name; inside an app, MLX Swift runs them on Mac, iPhone and iPad. Mind the memory: in our test a 4-bit 0.5B model peaked at about 0.36 GB on a Mac; larger models need far more.
- Apple's own: some are Core ML files ready for an app; others are research weights for PyTorch and need converting with coremltools first.
Reading the table
- Downloads are Hugging Face's count for the last 30 days, not all time. Hugging Face counts requests for a model's main file, so one install can count more than once.
- Size is the total of the repo's current files on Hugging Face, in the sizes it shows (1 GB = 1,000 MB). Some repos hold several variants of a model, so it can be larger than the one download you need; a dash means it has not been measured yet.
- Trending is Hugging Face's own score for recent attention. "from" under a name is the model it was made from.
- Parameters is the parameter count in the model files. A count marked * comes from the model's name (such as 27B) instead: the files gave none, or a much smaller one, as packed low-bit files can.
- Quantization is read from the model's tags and name, not checked against the files.
Before you use one
The list comes from Hugging Face's public API and is updated daily. We don't test each model. Check the license and read the model card before you ship one in an app: some licenses limit commercial use, and "other" means the terms are on the model card. Models whose names or tags point at adult or safety-removed content are left out, and so are gated models that need a Hugging Face sign-in to download.