Gemma 4 for iPhone

• Complete privacy — your data never leaves the device
• Works fully offline, no internet required
• Zero API costs — runs free once downloaded
• Sub-second response on A17 Pro / A18 chips

The E2B model is the top recommendation for iPhone. At ~4.6 GB download it fits within iPhone storage comfortably, and 4-bit quantization reduces active RAM to ~2.5 GB — well within the memory budget of iPhone 15/16 series.

The official path. Google AI Edge provides a native Swift package that runs Gemma 4 using Metal GPU acceleration.

• Add the Google AI Edge Swift package to your Xcode project
• Download the E2B model weights in TensorFlow Lite / LiteRT format
• Initialize the session and call generate()

Package.swift

// Package.swift
dependencies: [
    .package(
        url: "https://github.com/google/generative-ai-swift",
        from: "0.5.0"
    )
]

Swift Inference

import GoogleAIEdge

// Load E2B model (place .task file in app bundle)
let modelPath = Bundle.main.path(forResource: "gemma4-e2b-it-q4", ofType: "task")!
let session = try LlmInference(modelPath: modelPath)

// Generate response
let response = try await session.generateResponse(
    inputText: "Explain quantum computing in simple terms."
)
print(response)

For users who prefer a ready-made iOS app. LLM Farm and Offline Chat on the App Store both use llama.cpp under the hood and support Gemma GGUF models. Download the Q4_K_M quantized GGUF from Hugging Face, import it in the app, and start chatting. No code required.

Already running Ollama on a Mac? Expose the API on your local network and connect from the iPhone using any OpenAI-compatible app (e.g. Enchanted, OllamaChat). Set OLLAMA_HOST=0.0.0.0 on the Mac, then point the app to http://your-mac-ip:11434.

# On your Mac — allow LAN access
export OLLAMA_HOST=0.0.0.0
ollama serve

# Pull Gemma 4 E2B if not already done
ollama pull gemma4:e2b

# iPhone app settings:
# Base URL: http://192.168.x.x:11434  (your Mac's local IP)
# Model: gemma4:e2b

Older iPhones can run the E2B model via CPU inference but will be slow (2–4 tokens/sec). iPhone 15 Pro and later with the A17 Pro chip get Metal GPU acceleration and deliver practical real-time speed.

Tokens/sec measured with Gemma 4 E2B Q4_K_M, 512-token prompt, using Metal GPU. CPU-only fallback is 3–5× slower.

• Use Q4_K_M quantization — best quality/speed trade-off for mobile
• Close other apps before running inference to free RAM
• Keep prompts under 2K tokens for fastest response on phone
• Plug in charger during extended sessions — inference is GPU-intensive

• Gemma 4 Installation Guide — desktop / server setup
• Gemma 4 for Android — full Android guide
• Gemma 4 Local Deployment — advanced options