OpenMedKit Apple Platform Support¶

OpenMedKit supports macOS 14+, iOS 17+, watchOS 10+, and visionOS 1+. The runtime surface depends on the platform so constrained devices do not link backends that exceed their deployment or memory envelope.

The watchOS and visionOS limits use OpenMed's canonical Nano sub-tier: 10–30M parameters, at most 150 MB resident memory, and INT8 CoreML artifacts. These limits are enforced from model metadata before MLModel is opened. A Tiny, Base, over-budget, or non-INT8 model fails closed instead of being loaded.

Selecting and loading a constrained CoreML model¶

Describe the bundled model candidates and let PlatformModel select the highest-capacity candidate that fits the current target:

import OpenMedKit

let nano = PlatformModelDescriptor(
    identifier: "OpenMed-PII-Nano-INT8",
    modelURL: Bundle.main.url(
        forResource: "OpenMed-PII-Nano-INT8",
        withExtension: "mlmodelc"
    )!,
    id2labelURL: Bundle.main.url(
        forResource: "id2label",
        withExtension: "json"
    )!,
    tier: .nano,
    parameterCount: 24_000_000,
    estimatedResidentMemoryMB: 128,
    isINT8: true
)

let model = try PlatformModel(candidates: [nano])

watchOS and visionOS intentionally omit the full MLX and swift-transformers graph. Apps tokenize with the assets bundled beside their Nano model and pass bounded token IDs, attention masks, and character offsets to PlatformModel.predict(...). This keeps model and tokenizer access local; OpenMedKit does not add a cloud fallback.

Minimal redaction surface¶

PlatformModel.redact(...) applies mask or removal redaction to detected EntityPrediction spans without loading another backend:

let note = "Patient Ada Lovelace, MRN TEST-123."
let spans = [
    EntityPrediction(
        label: "full_name",
        text: "Ada Lovelace",
        confidence: 0.99,
        start: 8,
        end: 20
    ),
    EntityPrediction(
        label: "medical_record_number",
        text: "TEST-123",
        confidence: 0.99,
        start: 26,
        end: 34
    ),
]

let result = PlatformModel.redact(note, entities: spans)
// Patient [FULL_NAME], MRN [MEDICAL_RECORD_NUMBER].

Offsets remain character offsets into the original note. The watchOS and visionOS simulator tests use the same synthetic note and iOS reference spans, with a one-character tolerance at each boundary.

Build and validation¶

The Swift workflow performs the normal macOS tests, the iOS simulator build, and focused parity tests on available watchOS and visionOS simulators. Local checks use the same package scheme:

cd swift/OpenMedKit
xcodebuild build -scheme OpenMedKit \
  -destination 'generic/platform=watchOS Simulator'
xcodebuild build -scheme OpenMedKit \
  -destination 'generic/platform=visionOS Simulator'

Use only synthetic notes in committed tests and fixtures. OpenMedKit keeps inference on device and does not log input text or detected span text.