Build and run the default Android text-generation application

Connect an Android phone

Enable Developer options and USB debugging on the phone. Connect it with a data-capable USB cable, unlock it, and accept the debugging authorization prompt.

Verify the connection:

    

        
        
adb devices -l

    

The output must list the phone’s serial with the status device. Don’t continue if the list is empty or the status is unauthorized or offline. Unlock the phone and accept the debugging prompt if authorization is pending. Windows might also need the phone manufacturer’s USB driver. If the phone doesn’t appear, see Run apps on a hardware device .

Prepare the workspace and model downloader

Create the workspace used by the remaining pages:

    

        
        

mkdir -p "$HOME/text-to-text-android"
cd "$HOME/text-to-text-android"
python3 -m venv .hf-venv
.hf-venv/bin/python -m pip install --upgrade pip huggingface_hub
.hf-venv/bin/hf auth login
.hf-venv/bin/hf auth whoami
  

    

    

        
        

$TEXT_MODEL_WORKSPACE = Join-Path $env:USERPROFILE "text-to-text-android"
New-Item -ItemType Directory -Force -Path $TEXT_MODEL_WORKSPACE | Out-Null
Set-Location $TEXT_MODEL_WORKSPACE
python -m venv .hf-venv
.\.hf-venv\Scripts\python.exe -m pip install --upgrade pip huggingface_hub
.\.hf-venv\Scripts\hf.exe auth login
.\.hf-venv\Scripts\hf.exe auth whoami
  

    

The Arm AI Portal model packages that you’ll use are published by Arm and hosted in Arm’s Hugging Face repositories. Sign in with a Hugging Face account that has access to the model repositories and use a read token when prompted.

Download SmolLM2

Start with SmolLM2 360M Instruct through ExecuTorch. It’s the smallest validated text-generation example supplied with the Android application and provides the shortest path to a working result.

Run the command from the text-to-text-android workspace to download the complete optimized model package, including the .pte program, configuration, tokenizer, and chat template:

    

        
        

export MODEL_ID="smollm2-360m-instruct-8da4w-xnnpack-executorch"
.hf-venv/bin/hf download "Arm/$MODEL_ID" --local-dir ./model
  

    

    

        
        

$MODEL_ID = "smollm2-360m-instruct-8da4w-xnnpack-executorch"
.\.hf-venv\Scripts\hf.exe download "Arm/$MODEL_ID" --local-dir .\model
  

    

The model directory contains HuggingFaceTB__SmolLM2-360M-Instruct_executorch_optimized.pte, config.yaml, and the tokenizer directory. Keep the filenames and directory structure unchanged because the validated adapter checks those paths when it loads the model.

Model repositories can change over time. After downloading a model, compare the files in ./model with the model file expected by the selected validated example and with the inspection output.

Check the model package

The SmolLM2 ExecuTorch model card identifies the runtime, quantization, intended target, and required files. Before deploying another model, use its model card to confirm the same information, together with its Android and memory requirements.

The model package is a pre-exported runtime artifact. You don’t need to export, quantize, or tune the model.

Validated model examples

The starter application contains one validated adapter example for each model. Complete the SmolLM2 workflow first.

When you want to try another runtime or workload, return to the following table:

The model and runtime change together in these examples, so their results and timings aren’t controlled runtime benchmarks.

Clone the starter application

The starter application provides the shared model selector, text input, and Load and Run actions. The application also provides result display, model catalog, and an adapter interface.

Clone the starter application into the workspace:

    

        
        
git clone https://github.com/arm-education/ai-portal-android-app-text-to-text.git
cd ai-portal-android-app-text-to-text

    

Run the remaining commands from the repository root.

Apply the validated SmolLM2 adapter

The base project keeps its runtime dependencies and adapters minimal. Apply the supplied example that contains the SmolLM2 catalog entry, ExecuTorch adapter, and pinned ExecuTorch Android dependency:

    

        
        

cp -R adapter-examples/smollm2-executorch/app/. app/
  

    

    

        
        

Copy-Item -Path "adapter-examples\smollm2-executorch\app\*" `
    -Destination "app" -Recurse -Force
  

    

The example adds org.pytorch:executorch-android:1.1.0 to the build and registers the expected model and tokenizer files in model_catalog.json.

Build and install the application

Before building, confirm that Gradle is running with JDK 17:

    

        
        

./gradlew --version
  

    

    

        
        

.\gradlew.bat --version
  

    

The Launcher JVM line should report version 17. The Daemon JVM line should point to the JDK 17 installation or state that it uses the current Java home. If either JVM uses another version, set JAVA_HOME to your JDK 17 installation and stop the existing Gradle daemon before trying again:

    

        
        

./gradlew --stop
  

    

    

        
        

.\gradlew.bat --stop
  

    

Build and lint the debug APK:

    

        
        

chmod +x gradlew
./gradlew :app:assembleDebug :app:lintDebug
  

    

    

        
        

.\gradlew.bat :app:assembleDebug :app:lintDebug
  

    

Continue only when the Gradle output ends with BUILD SUCCESSFUL. A message that Gradle can’t strip a prebuilt .so library is a packaging warning and doesn’t cause the build to fail.

Confirm that the phone is still connected before installing the application:

    

        
        
adb devices -l

    

Continue only if the phone is listed with the status device. Then install the APK and start the main activity:

    

        
        
adb install -r app/build/outputs/apk/debug/app-debug.apk
adb shell am start -n com.arm.learningpath.texttotext/.ui.MainActivity

    

The application opens with SmolLM2 selected. Its status shows that the required model file is missing because the model package hasn’t been copied into the application’s private storage yet.

Understand the example files

The validated example changes only the model-specific parts of the application:

The downloaded model files remain outside the APK. This keeps the application build smaller, so you can replace the model package without embedding its weights in the application.

What you’ve accomplished and what’s next

You’ve now connected an Arm-based Android phone, downloaded the default model, applied its validated adapter, and installed an application ready to load SmolLM2.

Next, you’ll run the model on the Android phone.