Models
Core AI runs models exported with Apple’s coreai-models. Apple publishes the recipes, not the exported models, so we host ready-to-use ones. Your app downloads or bundles the models it needs and passes each folder as source.
Hosted Exports
Section titled “Hosted Exports”Tested on an iPhone 17 Pro with iOS 27 (Wan 2.1 on macOS) and hosted on Hugging Face. Download the folder from the Folder column into app storage, and point source at it. Each repository’s card lists its license.
| Task | Model | Hugging Face repository | Folder | Size | Platforms | Use with |
|---|---|---|---|---|---|---|
| Language | Qwen3 0.6B | grabbou/qwen3-0.6b-CoreAI |
ios/ |
456 MB | iOS | coreAI.languageModel(...) |
| Vision language | Qwen3-VL 2B | grabbou/Qwen3-VL-2B-Instruct-CoreAI |
qwen3_vl_2b/ |
2.9 GB | iOS (memory limit), macOS | coreAI.languageModel(...) |
| Image generation | FLUX.2 klein 4B | grabbou/FLUX.2-klein-4B-CoreAI |
ios/ |
4.0 GB | iOS | coreAI.imageModel(...) |
| Video generation | Wan 2.1 T2V 1.3B | grabbou/Wan2.1-T2V-1.3B-CoreAI |
ios/ |
4.3 GB | macOS (#306) | coreAI.videoModel(...) |
| Transcription | Parakeet TDT 0.6B v3 | grabbou/parakeet-tdt-0.6b-v3-CoreAI |
streaming/ |
1.3 GB | iOS, macOS | coreAI.transcriptionModel(...) |
| Embeddings | CLIP ViT-B/32 | grabbou/clip-vit-base-patch32-CoreAI |
clip-vit-base-patch32/ |
308 MB | iOS, macOS | coreAI.embeddingModel(...) |
| Object detection | YOLOS-tiny | grabbou/yolos-tiny-CoreAI |
yolos-tiny.aimodel/ |
24 MB | iOS, macOS | CoreAI.runTask('object-detection', ...) |
| Image segmentation | EfficientSAM-Ti | grabbou/EfficientSAM-Ti-CoreAI |
efficient-sam/ |
41 MB | iOS, macOS | CoreAI.runTask('segmentation', ...) |
| Video segmentation | SAM 3 (video) | grabbou/sam3-video-CoreAI |
sam3-video/ |
1.7 GB | iOS, macOS | CoreAI.runTask('video-segmentation', ...) |
| Super-resolution | EDSR x2 | grabbou/edsr-x2-CoreAI |
edsr-x2.aimodel/ |
2.8 MB | iOS, macOS | CoreAI.runTask('super-resolution', ...) |
| Classification | PVT v2 b0 (ImageNet-1k) | grabbou/pvt-v2-b0-CoreAI |
pvt-v2-b0/ |
7.4 MB | iOS, macOS | CoreAI.runTask('classification', ...) |
| Depth estimation | Depth Anything 3 small | grabbou/depth-anything-3-small-CoreAI |
da3-small.aimodel/ |
106 MB | iOS, macOS | CoreAI.runTask('depth', ...) |
Any .aimodel |
— | — | — | — | — | CoreAI.loadModel({ ..., task: 'raw' }), CoreAI.runRawFunction |
The weights keep their original licenses (Apache-2.0, MIT, CC BY 4.0, and the SAM License for SAM 3), which each card lists. The example app’s downloadHubFolder downloads one of these folders, and the harness tests show each model in use.
Community Exports
Section titled “Community Exports”Third-party uploads, not published by Apple or Callstack:
| Task | Model | Bundle | Platforms |
|---|---|---|---|
| Language | Qwen3 4B | iOS .aimodel, macOS .aimodel |
iOS, macOS |
| Language | gpt-oss-20b | macOS .aimodel |
macOS |
| Language | Gemma 3 4B it | macOS .aimodel |
macOS |
| Image segmentation | SAM 3 (image, text prompts) | segmenter bundle (1.7 GB) | iOS, macOS |
The package runs on iOS, so macOS-only models are listed for reference.
Making Your Own
Section titled “Making Your Own”For a model or size that isn’t hosted, export it yourself with Apple’s coreai-models recipes. The export guide covers setup, iOS presets and quantization, and each hosted repository’s card has the exact command used for it. Export for iOS where a recipe has a platform option.
To check what an export expects before wiring it to a task, list its functions with CoreAI.inspectModel.