Add multi-model download: Gemma 3n E4B (recommended), E2B, Gemma 2B

- ModelSpec enum with 3 downloadable options from MediaPipe CDN
- Gemma 3n E4B set as recommended default (best coding/reasoning via MoE)
- Gemma 3n E2B as faster/smaller alternative
- Gemma 2B kept as lightest option
- UI shows all 3 download buttons, hides all once any model is installed
- Installed model name shown in status (e.g. "✓ Gemma 3n E4B ready")
- Custom models (DeepSeek Coder, Qwen2.5-Coder) can be manually placed in files dir

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app/src/main/java/com/pixel10/ai/inference/ModelDownloader.kt

@@ -0,0 +1,145 @@
+package com.pixel10.ai.inference
+
+import android.content.Context
+import android.util.Log
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.withContext
+import java.io.File
+import java.io.FileOutputStream
+import java.net.HttpURLConnection
+import java.net.URL
+
+/**
+ * Downloads a MediaPipe-compatible model for background-safe inference.
+ *
+ * Gemini Nano (ML Kit) blocks inference when the app is backgrounded (ErrorCode 30).
+ * MediaPipe with a local model file has no such restriction — it runs entirely in
+ * the app process using the Tensor G5 GPU via OpenCL/Vulkan.
+ *
+ * Three model options (all from Google's MediaPipe CDN):
+ *  - [ModelSpec.GEMMA_3N_E4B_CODING]  — best coding/reasoning, ~2.5 GB (recommended)
+ *  - [ModelSpec.GEMMA_3N_E2B_CODING]  — good balance, ~1.5 GB
+ *  - [ModelSpec.GEMMA_2B_GENERAL]     — lightest, ~1.3 GB
+ *
+ * Custom models (DeepSeek Coder, Qwen2.5-Coder, etc.) can be placed manually in
+ * the app's files directory after converting with ai-edge-torch.
+ */
+object ModelDownloader {
+
+    private const val TAG = "ModelDownloader"
+
+    /** Available model specs that can be downloaded from Google's MediaPipe CDN. */
+    enum class ModelSpec(
+        val displayName: String,
+        val filename: String,
+        val url: String,
+        val sizeMb: Int,
+        val description: String
+    ) {
+        /** Recommended: best coding & reasoning quality via MoE architecture. */
+        GEMMA_3N_E4B_CODING(
+            displayName = "Gemma 3n E4B",
+            filename = "gemma-3n-E4B-it-int4.task",
+            url = "https://storage.googleapis.com/mediapipe-models/llm_inference/" +
+                  "gemma-3n-E4B-it-int4/float16/1/gemma-3n-E4B-it-int4.task",
+            sizeMb = 2500,
+            description = "Best coding & reasoning (~2.5 GB)"
+        ),
+        /** Good balance between quality and speed. */
+        GEMMA_3N_E2B_CODING(
+            displayName = "Gemma 3n E2B",
+            filename = "gemma-3n-E2B-it-int4.task",
+            url = "https://storage.googleapis.com/mediapipe-models/llm_inference/" +
+                  "gemma-3n-E2B-it-int4/float16/1/gemma-3n-E2B-it-int4.task",
+            sizeMb = 1500,
+            description = "Good balance, faster (~1.5 GB)"
+        ),
+        /** Lightest option — general-purpose, not optimised for code. */
+        GEMMA_2B_GENERAL(
+            displayName = "Gemma 2B",
+            filename = "gemma-2b-it-gpu-int4.bin",
+            url = "https://storage.googleapis.com/mediapipe-models/llm_inference/" +
+                  "gemma-2b-it-gpu-int4/float16/1/gemma-2b-it-gpu-int4.bin",
+            sizeMb = 1300,
+            description = "Lightest, general-purpose (~1.3 GB)"
+        )
+    }
+
+    data class Progress(
+        val downloadedBytes: Long,
+        val totalBytes: Long,
+        val percent: Int = if (totalBytes > 0) (downloadedBytes * 100 / totalBytes).toInt() else 0
+    )
+
+    /** Returns true if any supported model is present in the app's files directory. */
+    fun isModelPresent(context: Context): Boolean =
+        ModelSpec.values().any { modelFile(context, it).let { f -> f.exists() && f.length() > 1_000_000L } }
+
+    /** Returns the installed [ModelSpec], or null if no model is present. */
+    fun installedSpec(context: Context): ModelSpec? =
+        ModelSpec.values().firstOrNull { modelFile(context, it).let { f -> f.exists() && f.length() > 1_000_000L } }
+
+    fun modelFile(context: Context, spec: ModelSpec): File =
+        File(context.filesDir, spec.filename)
+
+    /** Legacy compat — returns the file of the installed model, or Gemma 3n E4B path as default. */
+    fun modelFile(context: Context): File =
+        installedSpec(context)?.let { modelFile(context, it) }
+            ?: modelFile(context, ModelSpec.GEMMA_3N_E4B_CODING)
+
+    /**
+     * Download [spec], reporting progress via [onProgress].
+     * Supports resume — if a partial file exists, continues from where it left off.
+     */
+    suspend fun download(
+        context: Context,
+        spec: ModelSpec = ModelSpec.GEMMA_3N_E4B_CODING,
+        onProgress: (Progress) -> Unit
+    ) = withContext(Dispatchers.IO) {
+        val dest = modelFile(context, spec)
+        val alreadyDownloaded = if (dest.exists()) dest.length() else 0L
+
+        Log.i(TAG, "Download starting ${spec.displayName} (already have $alreadyDownloaded bytes)")
+
+        val conn = URL(spec.url).openConnection() as HttpURLConnection
+        try {
+            conn.connectTimeout = 30_000
+            conn.readTimeout   = 60_000
+            if (alreadyDownloaded > 0) {
+                conn.setRequestProperty("Range", "bytes=$alreadyDownloaded-")
+            }
+            conn.connect()
+
+            val code = conn.responseCode
+            val resuming = code == HttpURLConnection.HTTP_PARTIAL   // 206
+            if (code != HttpURLConnection.HTTP_OK && !resuming) {
+                throw OnDeviceModel.InferenceException("Download failed: HTTP $code")
+            }
+
+            val serverBytes = conn.contentLengthLong.coerceAtLeast(0L)
+            val totalBytes  = if (resuming) alreadyDownloaded + serverBytes else serverBytes
+
+            conn.inputStream.use { input ->
+                FileOutputStream(dest, /* append= */ resuming).use { out ->
+                    val buf = ByteArray(128 * 1024)
+                    var written = alreadyDownloaded
+                    var read: Int
+                    while (input.read(buf).also { read = it } != -1) {
+                        out.write(buf, 0, read)
+                        written += read
+                        onProgress(Progress(written, totalBytes))
+                    }
+                }
+            }
+
+            Log.i(TAG, "Download complete — ${dest.length()} bytes")
+        } finally {
+            conn.disconnect()
+        }
+    }
+
+    fun deleteModel(context: Context) {
+        ModelSpec.values().forEach { modelFile(context, it).delete() }
+        Log.i(TAG, "All models deleted")
+    }
+}

app/src/main/java/com/pixel10/ai/inference/OnDeviceModel.kt

@@ -102,38 +102,43 @@ interface OnDeviceModel {
 
         /**
          * Create the best available on-device model.
-         * Tries Gemini Nano (Tensor G5 TPU) first, falls back to MediaPipe.
+         *
+         * Priority order:
+         *  1. MediaPipe (local model file) — runs in background, uses Tensor G5 GPU.
+         *     This is the preferred backend: no foreground restriction, no AICore dep.
+         *  2. Gemini Nano (ML Kit) — foreground only (ErrorCode 30 in background).
+         *     Used as fallback when no MediaPipe model file is present.
+         *
+         * Tap "Download Model" in the app UI to get the MediaPipe model automatically.
          */
         suspend fun create(context: Context): OnDeviceModel = withContext(Dispatchers.IO) {
-            // Try Gemini Nano via ML Kit Prompt API
+            // MediaPipe first — background-safe, GPU-accelerated via Tensor G5
             try {
-                Log.i(TAG, "Attempting Gemini Nano via ML Kit Prompt API...")
+                Log.i(TAG, "Attempting MediaPipe LLM with local model...")
+                val mediapipe = MediaPipeModel.create(context)
+                Log.i(TAG, "MediaPipe model ready: ${mediapipe.backendName}")
+                return@withContext mediapipe
+            } catch (e: Exception) {
+                Log.w(TAG, "MediaPipe not available: ${e.message}")
+            }
+
+            // Gemini Nano fallback — only works when app is in foreground
+            try {
+                Log.i(TAG, "Attempting Gemini Nano via ML Kit (foreground only)...")
                 val nano = GeminiNanoModel.create(context)
-                Log.i(TAG, "Gemini Nano ready!")
+                Log.i(TAG, "Gemini Nano ready (foreground only)")
                 return@withContext nano
             } catch (e: Exception) {
                 Log.w(TAG, "Gemini Nano not available: ${e.message}")
             }
 
-            // Fall back to MediaPipe with a local model file
-            try {
-                Log.i(TAG, "Attempting MediaPipe LLM with local model...")
-                val mediapipe = MediaPipeModel.create(context)
-                Log.i(TAG, "MediaPipe model ready!")
-                return@withContext mediapipe
-            } catch (e: Exception) {
-                Log.w(TAG, "MediaPipe model not available: ${e.message}")
-            }
-
             throw InferenceException(
-                "No on-device AI model available.\n\n" +
-                "Option 1: Use a Pixel device with Gemini Nano support " +
-                "(Pixel 10/9/8 series)\n\n" +
-                "Option 2: Place a MediaPipe-compatible model (.bin or .task) in:\n" +
+                "No model loaded yet.\n\n" +
+                "Tap 'Download Model' in the app to download Gemma 2B (~1.3 GB).\n" +
+                "Once downloaded the server works fully in the background.\n\n" +
+                "Or place a compatible model file in:\n" +
                 "  ${context.filesDir.absolutePath}/\n" +
-                "  Supported: gemma-3n-E2B.task, gemma-2b-it-gpu-int4.bin, etc.\n\n" +
-                "Download models from:\n" +
-                "  https://ai.google.dev/edge/mediapipe/solutions/genai/llm_inference/android"
+                "  Supported: gemma-2b-it-gpu-int4.bin, gemma-3n-E2B.task, etc."
             )
         }
     }

app/src/main/java/com/pixel10/ai/ui/MainActivity.kt

@@ -15,9 +15,13 @@ import android.view.View
 import androidx.activity.result.contract.ActivityResultContracts
 import androidx.appcompat.app.AppCompatActivity
 import androidx.core.content.ContextCompat
+import androidx.lifecycle.lifecycleScope
 import com.pixel10.ai.R
 import com.pixel10.ai.databinding.ActivityMainBinding
+import com.pixel10.ai.inference.ModelDownloader
+import com.pixel10.ai.inference.ModelDownloader.ModelSpec
 import com.pixel10.ai.server.ApiServerService
+import kotlinx.coroutines.launch
 import java.text.SimpleDateFormat
 import java.util.Date
 import java.util.Locale
@@ -27,6 +31,7 @@ class MainActivity : AppCompatActivity() {
     private lateinit var binding: ActivityMainBinding
     private var service: ApiServerService? = null
     private var bound = false
+    private var downloading = false
 
     private val logBuffer = StringBuilder()
 
@@ -66,20 +71,33 @@ class MainActivity : AppCompatActivity() {
         requestNotificationPermission()
 
         binding.btnToggle.setOnClickListener {
-            if (service?.isRunning == true) {
-                stopServer()
-            } else {
-                startServer()
-            }
+            if (service?.isRunning == true) stopServer() else startServer()
         }
 
+        binding.btnDownloadGemma3nE4b.setOnClickListener {
+            startModelDownload(ModelSpec.GEMMA_3N_E4B_CODING)
+        }
+        binding.btnDownloadGemma3nE2b.setOnClickListener {
+            startModelDownload(ModelSpec.GEMMA_3N_E2B_CODING)
+        }
+        binding.btnDownloadModel.setOnClickListener {
+            startModelDownload(ModelSpec.GEMMA_2B_GENERAL)
+        }
+
+        updateModelCard()
         updateStatus(ApiServerService.ServerState.STOPPED)
         appendLog("Pixel10 AI Server ready")
         appendLog("Device: ${Build.MANUFACTURER} ${Build.MODEL}")
         appendLog("SoC: ${Build.SOC_MODEL}")
         appendLog("Android ${Build.VERSION.RELEASE} (API ${Build.VERSION.SDK_INT})")
         appendLog("")
-        appendLog("Tap 'Start Server' to begin serving AI inference")
+        if (ModelDownloader.isModelPresent(this)) {
+            appendLog("Model ready — server works in background")
+        } else {
+            appendLog("No local model found")
+            appendLog("Tap 'Download Model' to enable background inference")
+            appendLog("(Without it, Gemini Nano only works in foreground)")
+        }
     }
 
     override fun onStart() {
@@ -110,16 +128,72 @@ class MainActivity : AppCompatActivity() {
         }
     }
 
+    private fun startModelDownload(spec: ModelSpec) {
+        if (downloading) return
+        downloading = true
+        setDownloadButtonsEnabled(false)
+        binding.progressDownload.visibility = View.VISIBLE
+        binding.tvModelDownloadStatus.text = "Starting download: ${spec.displayName}…"
+
+        lifecycleScope.launch {
+            try {
+                ModelDownloader.download(this@MainActivity, spec) { progress ->
+                    runOnUiThread {
+                        binding.progressDownload.progress = progress.percent
+                        val mb = progress.downloadedBytes / 1_048_576
+                        val total = progress.totalBytes / 1_048_576
+                        binding.tvModelDownloadStatus.text =
+                            "${spec.displayName}: ${mb}MB / ${total}MB (${progress.percent}%)"
+                    }
+                }
+                runOnUiThread {
+                    downloading = false
+                    updateModelCard()
+                    appendLog("${spec.displayName} downloaded — background inference enabled")
+                }
+            } catch (e: Exception) {
+                runOnUiThread {
+                    downloading = false
+                    setDownloadButtonsEnabled(true)
+                    binding.progressDownload.visibility = View.GONE
+                    binding.tvModelDownloadStatus.text = "Download failed: ${e.message}"
+                    appendLog("Download error: ${e.message}")
+                }
+            }
+        }
+    }
+
+    private fun setDownloadButtonsEnabled(enabled: Boolean) {
+        binding.btnDownloadGemma3nE4b.isEnabled = enabled
+        binding.btnDownloadGemma3nE2b.isEnabled = enabled
+        binding.btnDownloadModel.isEnabled = enabled
+    }
+
+    private fun updateModelCard() {
+        val spec = ModelDownloader.installedSpec(this)
+        if (spec != null) {
+            binding.tvModelDownloadStatus.text = getString(R.string.model_downloaded, spec.displayName)
+            binding.btnDownloadGemma3nE4b.visibility = View.GONE
+            binding.btnDownloadGemma3nE2b.visibility = View.GONE
+            binding.btnDownloadModel.visibility = View.GONE
+            binding.progressDownload.visibility = View.GONE
+        } else {
+            binding.tvModelDownloadStatus.text = getString(R.string.model_not_downloaded)
+            binding.btnDownloadGemma3nE4b.visibility = View.VISIBLE
+            binding.btnDownloadGemma3nE2b.visibility = View.VISIBLE
+            binding.btnDownloadModel.visibility = View.VISIBLE
+            setDownloadButtonsEnabled(true)
+            binding.progressDownload.visibility = View.GONE
+        }
+    }
+
     private fun startServer() {
         val port = binding.etPort.text.toString().toIntOrNull() ?: 8080
-
         val intent = Intent(this, ApiServerService::class.java).apply {
             action = ApiServerService.ACTION_START
             putExtra(ApiServerService.EXTRA_PORT, port)
         }
         startForegroundService(intent)
-
-        // Bind if not already bound
         if (!bound) {
             bindService(
                 Intent(this, ApiServerService::class.java),
@@ -127,7 +201,6 @@ class MainActivity : AppCompatActivity() {
                 Context.BIND_AUTO_CREATE
             )
         }
-
         updateStatus(ApiServerService.ServerState.LOADING_MODEL)
     }
 
@@ -137,6 +210,10 @@ class MainActivity : AppCompatActivity() {
     }
 
     private fun updateStatus(state: ApiServerService.ServerState) {
+        val canEdit = state == ApiServerService.ServerState.STOPPED ||
+                      state == ApiServerService.ServerState.ERROR
+        binding.etPort.isEnabled = canEdit
+
         when (state) {
             ApiServerService.ServerState.STOPPED -> {
                 binding.tvServerStatus.text = getString(R.string.server_status_stopped)
@@ -145,14 +222,12 @@ class MainActivity : AppCompatActivity() {
                 binding.tvModelStatus.text = "Model: not loaded"
                 binding.btnToggle.text = getString(R.string.btn_start)
                 binding.btnToggle.isEnabled = true
-                binding.etPort.isEnabled = true
             }
             ApiServerService.ServerState.LOADING_MODEL -> {
                 binding.tvServerStatus.text = getString(R.string.server_status_starting)
                 (binding.viewStatusDot.background as? GradientDrawable)?.setColor(getColor(R.color.primary))
                 binding.tvModelStatus.text = getString(R.string.model_loading)
                 binding.btnToggle.isEnabled = false
-                binding.etPort.isEnabled = false
             }
             ApiServerService.ServerState.RUNNING -> {
                 val port = binding.etPort.text.toString()
@@ -163,7 +238,6 @@ class MainActivity : AppCompatActivity() {
                 binding.tvModelStatus.text = getString(R.string.model_ready)
                 binding.btnToggle.text = getString(R.string.btn_stop)
                 binding.btnToggle.isEnabled = true
-                binding.etPort.isEnabled = false
             }
             ApiServerService.ServerState.ERROR -> {
                 binding.tvServerStatus.text = getString(R.string.server_status_error)
@@ -171,7 +245,6 @@ class MainActivity : AppCompatActivity() {
                 binding.tvModelStatus.text = getString(R.string.model_error)
                 binding.btnToggle.text = getString(R.string.btn_start)
                 binding.btnToggle.isEnabled = true
-                binding.etPort.isEnabled = true
             }
         }
     }
@@ -180,16 +253,8 @@ class MainActivity : AppCompatActivity() {
         val timestamp = SimpleDateFormat("HH:mm:ss", Locale.US).format(Date())
         logBuffer.append("[$timestamp] $message\n")
         binding.tvLog.text = logBuffer.toString()
-
-        // Auto-scroll to bottom
-        binding.scrollLog.post {
-            binding.scrollLog.fullScroll(View.FOCUS_DOWN)
-        }
-
-        // Update request count
-        service?.let {
-            binding.tvRequestCount.text = "Requests served: ${it.requestCount}"
-        }
+        binding.scrollLog.post { binding.scrollLog.fullScroll(View.FOCUS_DOWN) }
+        service?.let { binding.tvRequestCount.text = "Requests served: ${it.requestCount}" }
     }
 
     @Suppress("DEPRECATION")
@@ -202,7 +267,6 @@ class MainActivity : AppCompatActivity() {
             }
         } catch (_: Exception) {}
 
-        // Fallback: iterate network interfaces
         try {
             val interfaces = java.net.NetworkInterface.getNetworkInterfaces()
             while (interfaces.hasMoreElements()) {

app/src/main/res/layout/activity_main.xml

@@ -105,6 +105,74 @@
         </LinearLayout>
     </com.google.android.material.card.MaterialCardView>
 
+    <!-- Model Download Card -->
+    <com.google.android.material.card.MaterialCardView
+        android:id="@+id/cardModel"
+        android:layout_width="0dp"
+        android:layout_height="wrap_content"
+        android:layout_marginTop="12dp"
+        app:cardBackgroundColor="@color/surface_variant"
+        app:cardCornerRadius="16dp"
+        app:cardElevation="0dp"
+        app:strokeWidth="0dp"
+        app:layout_constraintTop_toBottomOf="@id/cardStatus"
+        app:layout_constraintStart_toStartOf="parent"
+        app:layout_constraintEnd_toEndOf="parent">
+
+        <LinearLayout
+            android:layout_width="match_parent"
+            android:layout_height="wrap_content"
+            android:orientation="vertical"
+            android:padding="16dp">
+
+            <TextView
+                android:id="@+id/tvModelDownloadStatus"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:textColor="@color/log_text"
+                android:textSize="13sp" />
+
+            <ProgressBar
+                android:id="@+id/progressDownload"
+                style="@android:style/Widget.ProgressBar.Horizontal"
+                android:layout_width="match_parent"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="8dp"
+                android:max="100"
+                android:visibility="gone" />
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnDownloadGemma3nE4b"
+                style="@style/Widget.MaterialComponents.Button.OutlinedButton"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="8dp"
+                android:text="@string/btn_download_gemma3n_e4b"
+                android:textSize="13sp"
+                app:cornerRadius="8dp" />
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnDownloadGemma3nE2b"
+                style="@style/Widget.MaterialComponents.Button.OutlinedButton"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="4dp"
+                android:text="@string/btn_download_gemma3n_e2b"
+                android:textSize="13sp"
+                app:cornerRadius="8dp" />
+
+            <com.google.android.material.button.MaterialButton
+                android:id="@+id/btnDownloadModel"
+                style="@style/Widget.MaterialComponents.Button.OutlinedButton"
+                android:layout_width="wrap_content"
+                android:layout_height="wrap_content"
+                android:layout_marginTop="4dp"
+                android:text="@string/btn_download_model"
+                android:textSize="13sp"
+                app:cornerRadius="8dp" />
+        </LinearLayout>
+    </com.google.android.material.card.MaterialCardView>
+
     <!-- Port Config -->
     <LinearLayout
         android:id="@+id/layoutPort"
@@ -113,7 +181,7 @@
         android:orientation="horizontal"
         android:gravity="center_vertical"
         android:layout_marginTop="16dp"
-        app:layout_constraintTop_toBottomOf="@id/cardStatus"
+        app:layout_constraintTop_toBottomOf="@id/cardModel"
         app:layout_constraintStart_toStartOf="parent"
         app:layout_constraintEnd_toEndOf="parent">
 

app/src/main/res/values/strings.xml

@@ -18,6 +18,11 @@
     <string name="model_not_loaded">Model: not loaded</string>
 
     <!-- Controls -->
+    <string name="btn_download_gemma3n_e4b">⭐ Gemma 3n E4B — Best coding (~2.5 GB)</string>
+    <string name="btn_download_gemma3n_e2b">Gemma 3n E2B — Faster (~1.5 GB)</string>
+    <string name="btn_download_model">Gemma 2B — Lightest (~1.3 GB)</string>
+    <string name="model_downloaded">✓ %s ready — background inference enabled</string>
+    <string name="model_not_downloaded">No local model. Download one to enable background inference.</string>
     <string name="btn_start">Start Server</string>
     <string name="btn_stop">Stop Server</string>
     <string name="port_label">Port:</string>