GitHub - vincentamato/MLXDINOv3: A native Swift implementation of Meta’s DINOv3 using MLX Swift.

MLXDINOv3

A native Swift implementation of Meta’s DINOv3 using MLX Swift.

DINOv3 is a family of self-supervised vision foundation models from Meta AI, producing high-quality dense visual features that outperform specialized models without fine-tuning. This package provides a numerically validated, on-device compatible version for Apple silicon.

Installation

Add MLXDINOv3 to your Swift Package Manager dependencies:

dependencies: [
    .package(url: "https://github.com/vincentamato/MLXDINOv3.git", from: "1.0.0")
]

Then import it:

Converting Hugging Face weights to MLX format

Convert Hugging Face weights to MLX format using the conversion CLI in Xcode:

1. Open the package in Xcode: xed .
2. Select the Convert scheme from the scheme selector
3. Edit the scheme (Product → Scheme → Edit Scheme)
4. Under "Run" → "Arguments", add:
   • facebook/dinov3-vits16-pretrain-lvd1689m
   • ./Models/dinov3-vits16-mlx
5. Run the scheme (Cmd+R)

Note: Currently, only the ViT models are supported.

Example Usage

import AppKit
import MLX
import MLXDINOv3

// Load a pretrained model
let model = try loadPretrained(modelPath: "Models/dinov3-vits16-mlx")

// Preprocess an image
let image = NSImage(contentsOfFile: "image.jpg")!
let processor = ImageProcessor()
let inputs = try processor(image)

// Run inference
let outputs = model(inputs)

print("Pooler output shape:", outputs.poolerOutput.shape)
print("Last hidden state shape:", outputs.lastHiddenState.shape)

Testing

All testing must be done from Xcode due to MLX metallib requirements.

Step 1: Convert the test model

1. Open the package in Xcode: xed .
2. Select the Convert scheme
3. Edit the scheme (Product → Scheme → Edit Scheme)
4. Under "Run" → "Arguments", add:
   • facebook/dinov3-vits16-pretrain-lvd1689m
   • Tests/MLXDINOv3Tests/Resources/Model
5. Run (Cmd+R)

Step 2: Run tests

Run tests with Cmd+U or Product → Test

Tests automatically download PyTorch reference outputs from HuggingFace Hub for validation.

References

• DINOv3 Paper: DINOv3
• DINOv3 Repository: facebookresearch/dinov3

License

This package is released under the MIT License.

Note: The pretrained DINOv3 weights and original model architecture are released under Meta’s DINOv3 License.