Palm detector — finds hands and seven detector keypoints.
Hand landmark detector — predicts presence, handedness, and 21 three-dimensional landmarks from each detected hand crop.

Together they enable gesture-controlled apps, games, plugins, and agents. Gesture recognition itself is implemented by interpreting the 21 landmarks; these models do not contain a separate gesture classifier. See Hugging Mac for the complete local macOS pipeline.

Model details

Component	Input	Outputs	Parameters
`hand_detector.mlpackage`	`image`: `1 × 3 × 256 × 256` FP32 RGB	`box_coords`: `1 × 2944 × 18`; `box_scores`: `1 × 2944 × 1`	1.76M
`hand_landmark_detector.mlpackage`	`image`: `1 × 3 × 256 × 256` FP32 RGB crop	`scores`: `1`; `lr`: `1`; `landmarks`: `1 × 21 × 3`	2.01M

Precision: FP32
Core ML deployment target: macOS 13 or later
Combined package size: 15.3 MB
The landmark model can be loaded lazily only when landmarks are requested.

Core ML example

python

from pathlib import Path

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

root = Path(snapshot_download(
    repo_id="hugging-mac/mediapipe-hand-coreml",
    allow_patterns=["coreml/**"],
)) / "coreml"

detector = ct.models.MLModel(
    root / "hand_detector.mlpackage",
    compute_units=ct.ComputeUnit.ALL,
)
image = ImageOps.pad(Image.open("hand.jpg").convert("RGB"), (256, 256))
value = np.asarray(image, dtype=np.float32) / 255.0
tensor = np.ascontiguousarray(value.transpose(2, 0, 1))[None]
outputs = detector.predict({"image": tensor})

print(outputs["box_coords"].shape)  # (1, 2944, 18)
print(outputs["box_scores"].shape)  # (1, 2944, 1)

Raw detector outputs require MediaPipe anchor decoding and weighted NMS. The resulting rotated hand crop is passed to the landmark model. Use the Hugging Mac MediaPipe Hand SDK for the complete pipeline, coordinate projection, and handedness handling.

Conversion verification

The Core ML models were compared with ONNX Runtime on identical random inputs. Maximum absolute differences:

Detector boxes: 6.866e-5
Detector scores: 1.526e-4
Hand presence: 3.92e-8
Handedness: 4.59e-6
Landmarks: 4.47e-7

Provenance and integrity

Upstream: Qualcomm MediaPipe Hand Detection
Upstream revision: 013e27b599e37c3b4c69439de15de53cc5b5708e
Qualcomm AI Hub release: v0.60.0, ONNX float
hand_detector.mlpackage SHA-256: a81143ef0bfa896d4206caf226697141ab8c052d57ba5e95fc7c7ac2660dc83d
hand_landmark_detector.mlpackage SHA-256: d4049538d8fe4b4ffa77f71e1efa43705fc010e721c0456a95d9c97453c13716
Complete bundle SHA-256: 5432c988be7ed2fd6847097f823e0d813f97be7dc7a7d6587129386dbcd75929

License

The upstream repository declares a custom license and links to the MediaPipePyTorch license, which states that the implementation follows the Apache License 2.0 terms used by MediaPipe. Review the upstream Qualcomm distribution terms and linked source license before redistribution or commercial use. Hugging Mac is not affiliated with or endorsed by Qualcomm, Google, or MediaPipe.