Argmax Local Server is now available!
Docs
Managing Models

Managing Models

Picking, downloading, and compiling models in your application

Downloading and Compiling Models

Argmax Pro SDK provides utilities to streamline model downloads and on-device compilation in your application. There are two main approaches:

  • ModelStore Approach (Recommended)
  • Direct Initializer Approach (Legacy)

ModelStore Approach

Key Benefits

  • Progress Tracking: Real-time download progress updates
  • Cancellation Support: Ability to cancel downloads
  • Repository Management: Utility to manage models across multiple model repositories
  • Decoupled model download and compilation: On-device model compilation (Specialization) happens separately from download, reducing blocking time

Basic Usage

@available(iOS 17.0, macOS 14.0, watchOS 10.0, visionOS 1.0, *)
public func setupWhisperKitPro() async throws -> WhisperKitPro {
 
    // First, obfuscate your Argmax API Key using `ObfuscatedKeyProvider.generateCodeExample`
    // `ObfuscatedKeyProvider` protects against casual inspection and static analysis tools.
    // In production, consider fetching apiKey securely from your backend.
    var keyProvider = ObfuscatedKeyProvider(mask: 37)
    keyProvider.apiKeyObfuscated = [4, 5, 6]
    
    guard let apiKey = keyProvider.apiKey else {
        fatalError("Missing API key")
    }
 
    // Initialize Argmax SDK to ensure Pro SDK licensed
    // This requires a connection during first use and once every 30 days
    await ArgmaxSDK.with(ArgmaxConfig(apiKey: apiKey))
 
    let modelRepo = RepoType.parakeetRepo  // e.g. RepoType.openSourceRepo, RepoType.proRepo
    
    // Fetch the recommended model for the current device
    let model = await WhisperKit.recommendedRemoteModels(
        from: modelRepo.rawValue,
    ).default
    
    let config = WhisperKitProConfig(
        model: model,
        modelRepo: modelRepo.rawValue,
    )
    let modelStore = ModelStore(whisperKitConfig: config)
    
    // Download model with progress tracking, print the progress
    // Or track the @Published `progress` in SwiftUI
    let cancellable = modelStore.$progress.sink { progress in
        if let progress = progress {
            let percentage = Int(progress.fractionCompleted * 100)
            print("\rDownload progress: \(percentage)%", terminator: "")
        }
    }
    
    // Download model (blocks until complete, track progress via modelStore.$progress)
    let modelURL = try await modelStore.downloadModel() 
 
    // Load model (blocks until complete, also compiles the model during first use or after OS update)
    let whisperKitPro = try await WhisperKitPro(config) 
    return whisperKitPro
}

Migrate from Direct Initializer Approach

If you're currently using the direct initializer approach, here's how to migrate:

Before (Direct Initializer)

// Old approach - blocks for entire process (download + compilation)
let config = WhisperKitProConfig(model: "large-v3-v20240930_626MB")
let whisperKitPro = try await WhisperKitPro(config)

After (ModelStore)

let config = WhisperKitProConfig(
    model: "large-v3-v20240930_626MB",
    modelRepo: RepoType.openSourceRepo.rawValue,
)
let modelStore = ModelStore(whisperKitConfig: config)
 
// Track progress and allow cancellation
let cancellable = modelStore.$progress.sink { progress in
    // Update UI with progress
}
 
let modelURL = try await modelStore.downloadModel()
 
// Initialize with downloaded model
let config = WhisperKitProConfig(
    modelFolder: modelURL.path(percentEncoded: false)
)
let whisperKitPro = try await WhisperKitPro(config)

Picking the Right Model

Argmax Pro SDK works with several model repositories out-of-the-box. This section shows you when and how to switch between them:

Nvidia Parakeet

This model is 9x faster than Whisper Large v3 Turbo on English speech-to-text and achieves slightly higher accuracy.

We recommend using this model for all applications that are English-only. Here is the diff to upgrade to this model:

let downloadURL = try await modelStore.downloadModel(
-    name: "large-v3-v20240930_626MB",
-    repo: .openSourceRepo
+    name: "parakeet-v2_476MB",
+    repo: .parakeetRepo
)

Nvidia Canary

Coming in late August 2025.

OpenAI Whisper

.proRepo hosts Whisper models that are further optimized for speed and energy-efficiency compared to their .openSourceRepo counterparts. During this upgrade, accuracy remains identical while speed and energy-efficiency improves significantly.

In order to use the upgraded Whisper models, simply switch to .proRepo in your initial configuration code:

let downloadURL = try await modelStore.downloadModel(
    name: "large-v3-v20240930_626MB",
-    repo: .openSourceRepo
+    repo: .proRepo
)
Managing Free and Paid UsersFile Transcription