Prepare models for offline use.

Lifecycle

let llm = PrimordialClient().localLLM()

try await llm.download()
try await llm.load()

try await llm.downloadAndLoad()

Storage Options

The default check requires enough space for remaining download bytes plus a safety margin. Customize or disable it with PrimordialDownloadOptions.

let options = PrimordialDownloadOptions(
    storageSafetyMarginFraction: 0.15,
    minimumStorageSafetyMarginBytes: 512 * 1024 * 1024
)

try await llm.download(options: options)
try await llm.download(options: .withoutStorageCheck)

Progress

try await llm.download { progress in
    print(progress.fractionCompleted)
    print(progress.status)
    print(progress.currentFilename ?? "")
    print(progress.downloadedBytes)
    print(progress.totalBytes ?? 0)
    print(progress.currentFileDownloadedBytes)
    print(progress.currentFileTotalBytes ?? 0)
    print(progress.bytesPerSecond ?? 0)
}

totalBytes and currentFileTotalBytes are optional. Use an indeterminate UI when totals are unavailable.

Cancel, Resume, Cleanup

let task = Task {
    try await llm.download()
}

task.cancel()

let downloaded = await llm.isDownloaded
let bytes = try await llm.downloadedModelSize()

try await llm.removeDownloadedModel()

Starting download() again reuses valid cached state and resumes supported partial transfers.

Voice Models

let models = PrimordialClient().voiceModelManager()

try await models.download { progress in
    print(progress.fractionCompleted)
    print(progress.phase)
}

try await models.load()

Voice progress is coarser than local LLM progress and does not currently include filename, byte totals, per-file bytes, or transfer speed.