---
license: agpl-3.0
library_name: coremltools
pipeline_tag: object-detection
base_model: Ultralytics/YOLOv8
tags:
  - coreml
  - apple-silicon
  - yolo
  - object-detection
  - macos
---

# YOLOv8 Object Detection — Core ML

Ready-to-use Core ML conversions of Ultralytics YOLOv8 for object detection on Apple Silicon. The `n`, `s`, and `m` variants are provided as `.mlpackage` models for use with the CPU, GPU, and Apple Neural Engine.

These models are used by [Hugging Mac](https://github.com/devilyouwei/hugging-mac), an open-source platform for building and experiencing local AI apps, services, games, plugins, and agents on macOS.

## Models

| Variant | Parameters | Package size | Best for |
|---|---:|---:|---|
| `yolov8n` | 3.2M | 6.5 MB | Lowest latency and resource usage |
| `yolov8s` | 11.2M | 22.5 MB | Balanced speed and accuracy |
| `yolov8m` | 25.9M | 52.0 MB | Higher accuracy |

## Provenance and conversion

- Source: [Ultralytics/YOLOv8](https://huggingface.co/Ultralytics/YOLOv8)
- Source revision: `8a9e1a5`
- Task: COCO 80-class object detection
- Input size: fixed `640 × 640`, batch size 1
- Precision: FP16 conversion
- NMS: not embedded; apply confidence filtering and NMS after inference

## Input and output

| Name | Type | Shape | Description |
|---|---|---|---|
| `image` | RGB image | `640 × 640` | Letterboxed model input |
| `predictions` | FP32 multi-array | `1 × 84 × 8400` | Boxes and scores for 80 COCO classes |

The first four output channels contain bounding boxes in `x, y, width, height` form. The remaining 80 channels contain class scores. Resize boxes back to the original image after confidence filtering and class-aware NMS.

## Core ML example

```python
from pathlib import Path

import coremltools as ct
from huggingface_hub import snapshot_download
from PIL import Image, ImageOps

root = Path(snapshot_download(
    repo_id="hugging-mac/yolov8-coreml",
    allow_patterns=["yolov8n.mlpackage/**"],
))

model = ct.models.MLModel(
    root / "yolov8n.mlpackage",
    compute_units=ct.ComputeUnit.ALL,
)

image = Image.open("image.jpg").convert("RGB")
image = ImageOps.pad(image, (640, 640), color=(114, 114, 114))
predictions = model.predict({"image": image})["predictions"]

print(predictions.shape)  # (1, 84, 8400)
```

For a complete preprocessing, decoding, and NMS implementation, see the [Hugging Mac YOLOv8 SDK](https://github.com/devilyouwei/hugging-mac/tree/main/packages/hugging_mac_sdk/src/hugging_mac_sdk/models/yolov8).

## Integrity

| Package | Directory SHA-256 |
|---|---|
| `yolov8n.mlpackage` | `c8eb62862dddac01512e47469e5dc15ad707d85c2d713d78a5b17b17d365f4a4` |
| `yolov8s.mlpackage` | `c5178bbaf83f8d4be513e2ce6f8081d44882466ab4a4eb62c704e1c9c95d3eac` |
| `yolov8m.mlpackage` | `d4e4e1a92b891cd8cfc3c41e15a57b92b26abc263deed6dc409874c1fbefe6a5` |

## License

The converted models retain the upstream [Ultralytics YOLOv8](https://github.com/ultralytics/ultralytics) licensing terms. They are published under AGPL-3.0. Review Ultralytics licensing requirements before commercial or closed-source use.

Hugging Mac is an independent open-source project and is not affiliated with or endorsed by Ultralytics.