Model	Params	Size	Input	Embedding	Accuracy
PE-Core-G14-448-ANE	2.4B	3.5GB	448x448	1280	1.0000
PE-Core-L-14-336-ANE	300M	604MB	336x336	1024	1.0000
PE-Core-B-16-ANE	86M	178MB	224x224	768	0.9998
PE-Core-S-16-384-ANE	22M	45MB	384x384	384	1.0000
PE-Core-T-16-384-ANE	6M	12MB	384x384	192	0.9999

Performance (M3 Mac)

Model	ANE Latency	MPS Latency	Speedup
PE-Core-bigG-14-448	783ms	1049ms	1.34x
PE-Core-L-14-336	~180ms	~280ms	~1.5x
PE-Core-B-16	~50ms	~80ms	~1.6x

Usage (Python)

python

import coremltools as ct
import numpy as np

# Load model
model = ct.models.MLModel("PE-Core-B-16-ANE.mlpackage")

# Prepare image (1, 3, 224, 224) normalized
image = np.random.randn(1, 3, 224, 224).astype(np.float32)

# Get embedding
output = model.predict({"image": image})
embedding = output["embedding"]  # (1, 768)

# Normalize for similarity search
embedding = embedding / np.linalg.norm(embedding)

Usage (Swift)

swift

import CoreML

let model = try MLModel(contentsOf: modelURL)
let input = try MLDictionaryFeatureProvider(dictionary: ["image": pixelBuffer])
let output = try model.prediction(from: input)
let embedding = output.featureValue(for: "embedding")!.multiArrayValue!

Conversion Details

Source: Meta's Perception Encoder via open_clip
Format: CoreML mlpackage (FP16)
Target: macOS 14+ (ANE optimized)
Accuracy: >99.98% cosine similarity vs PyTorch

Credits

Original models: Meta AI Perception Encoder
Loaded via: open_clip
Converted with: coremltools

SashimiSaketoro/PE-Core-ANE

PE-Core ANE (Apple Neural Engine) Models

Models

Performance (M3 Mac)

Usage (Python)

Usage (Swift)

Conversion Details

Credits

SashimiSaketoro/
PE-Core-ANE