# Kuzco - On-Device AI SDK for iOS

> Kuzco is a Swift SDK that enables iOS developers to run LLMs, vision models, and image generation locally on iPhone and iPad. No API fees, no internet required, complete user privacy.

Kuzco eliminates per-token API costs by running AI models directly on Apple devices. Developers pay a single flat monthly fee and can offer unlimited AI access to their users. The SDK supports text generation (Qwen, LLaMA, Phi, Gemma, DeepSeek), vision AI (image analysis, OCR, visual Q&A), and image generation (Stable Diffusion 2.1).

Key features:
- 3 lines of Swift code to integrate
- Streaming and one-shot text generation
- Multiple model options optimized for different use cases
- SwiftUI ready with async/await support
- Complete privacy - all processing happens on-device
- No internet connection required after model download

## Documentation

- [Getting Started](https://kuzco.co/docs): Overview and quick links to all documentation
- [Installation Guide](https://kuzco.co/docs/installation): Swift Package Manager setup and API key initialization
- [Quick Start](https://kuzco.co/docs/quick-start): Build your first AI feature in minutes
- [Available Models](https://kuzco.co/docs/models): All text, vision, and image generation models
- [Text Generation](https://kuzco.co/docs/text-generation): Streaming responses and chat interfaces
- [Image Generation](https://kuzco.co/docs/image-generation): Stable Diffusion on iOS
- [Vision AI](https://kuzco.co/docs/vision): Image analysis and visual question answering
- [Configuration](https://kuzco.co/docs/configuration): Temperature, tokens, and presets
- [Model Management](https://kuzco.co/docs/model-management): Download and storage management
- [Custom API](https://kuzco.co/docs/custom-api): Self-hosted model integration
- [API Reference](https://kuzco.co/docs/api-reference): Complete Swift types and methods

## Models

Text Models:
- Qwen3 4B (~2.5 GB, 32K context) - Best balance of quality and speed
- Qwen3 8B (~5 GB, 32K context) - Higher quality responses
- LLaMA 3 3B (~2 GB, 8K context) - Fastest generation
- Phi-4 Mini (~2.3 GB, 16K context) - Strong at coding
- Gemma 3 4B (~2.7 GB, 8K context) - Optimized for mobile
- DeepSeek R1 1.5B (~1 GB, 4K context) - Smallest footprint

Vision Models:
- Qwen3 VL (~4 GB) - Full-featured image understanding
- SmolVLM (~2 GB) - Compact vision-language model

Image Generation:
- Stable Diffusion 2.1 (~3.5 GB) - Text-to-image generation

## Code Example

```swift
import Kuzco

// Initialize once at app startup
KuzcoClient.initialize(apiKey: "kzc_your_api_key")

// Create a session and stream responses
let session = try await KuzcoSession(model: .qwen3_4b)

for try await partial in session.streamResponse(to: "Hello!") {
    print(partial.text, terminator: "")
}
```

## Optional

- [Join Waitlist](https://kuzco.co/waitlist): Sign up for early access
- [Home Page](https://kuzco.co): Main website