Gemma 4 for Android

• Full privacy — prompts and responses never leave your device
• Works offline — no Wi-Fi or mobile data needed
• Free after download — no API bills, no subscriptions
• Instant response on Snapdragon 8 Gen 3 / Dimensity 9300

E2B Q4_K_M is the recommended starting point for Android. It fits in ~2.4 GB of storage, uses ~2.5 GB RAM during inference, and runs at practical speeds on any flagship chipset from 2023 onward.

The official Google path. The AI Edge SDK integrates directly with Android Studio and uses GPU delegate for hardware-accelerated inference. Ideal for building your own app.

build.gradle

// build.gradle (app)
dependencies {
    implementation("com.google.ai.edge.litert:litert:1.0.1")
    implementation("com.google.ai.edge.litert:litert-gpu:1.0.1")
}

Kotlin Inference

import com.google.ai.edge.litert.LiteRtSession

// Initialize with E2B model asset
val session = LiteRtSession.createFromAsset(
    context,
    "gemma4-e2b-it-q4.task",
    LiteRtSession.Options.builder()
        .setPreferredBackend(LiteRtSession.Backend.GPU)
        .build()
)

// Generate response
val result = session.generateResponse("What is Gemma 4?")
println(result)

MLC-LLM provides a pre-built Android APK that runs Gemma 4 GGUF models using Vulkan GPU acceleration. Download the APK from the MLC-LLM GitHub releases, install it, then load the Gemma 4 E2B Q4 model from within the app. No coding required — works on any Android 10+ device with a modern GPU.

For developers who want full control. Install Termux from F-Droid, compile llama.cpp with OpenCL or Vulkan support, download the GGUF model, and run inference from the terminal. Slower setup but gives you direct access to all llama.cpp flags including context length, temperature, and batch size.

# Install Termux from F-Droid (not Play Store)
# Then inside Termux:
pkg update && pkg install clang cmake git

# Clone and build llama.cpp with Vulkan
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release -j4

# Download Gemma 4 E2B Q4_K_M GGUF
# (copy to phone storage or wget from Hugging Face)

# Run inference
./build/bin/llama-cli \
  -m /sdcard/gemma4-e2b-q4_k_m.gguf \
  -p "Explain machine learning simply" \
  -n 200 --gpu-layers 99

Mid-range chipsets (Snapdragon 7s Gen 2 and below) can run E2B Q4 via CPU but expect 2–4 tokens/sec. Flagship chips with Vulkan/OpenCL GPU compute deliver 10–20+ tokens/sec.

Tokens/sec measured with Gemma 4 E2B Q4_K_M, 512-token prompt. GPU path uses Vulkan compute via MLC-LLM or AI Edge SDK.

• Enable GPU acceleration — Vulkan delegate is 4–6× faster than CPU
• Use Q4_K_M quantization for best quality/speed on mobile
• Set Android performance mode (Settings → Battery → Performance) before long sessions
• Keep prompts under 2K tokens — longer contexts slow first-token latency significantly

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