expo-ai-kit 0.2.1 → 0.3.1

package/README.md

@@ -2,6 +2,8 @@
 
 On-device AI for Expo apps. Run language models locally—no API keys, no cloud, just native intelligence.
 
+**Now with Gemma 4 support** — Download and run Google's [Gemma 4](https://blog.google/technology/developers/gemma-4/) E2B (2.3B) and E4B (4.5B) models directly on Android devices via [LiteRT-LM](https://ai.google.dev/edge/litert-lm). Full on-device inference with GPU acceleration, streaming, and zero cloud dependency.
+
 [![npm version](https://img.shields.io/npm/v/expo-ai-kit.svg)](https://www.npmjs.com/package/expo-ai-kit)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 
@@ -14,6 +16,15 @@ On-device AI for Expo apps. Run language models locally—no API keys, no cloud,
 | iOS 26+ | [Apple Foundation Models](https://developer.apple.com/documentation/FoundationModels) |
 | Android (supported devices) | [ML Kit Prompt API](https://developers.google.com/ml-kit/genai#prompt-device) |
 
+### Downloadable Models (Gemma 4)
+
+| Platform | Status |
+|----------|--------|
+| Android | Gemma 4 E2B (2.3B) and E4B (4.5B) via [LiteRT-LM](https://ai.google.dev/edge/litert-lm) |
+| iOS | Coming soon — waiting for LiteRT-LM Swift APIs from Google |
+
+> **Note:** iOS downloadable model support (Gemma 4 E2B/E4B) is planned for a future release. We are waiting for Google to ship native Swift APIs for LiteRT-LM. Built-in Apple Foundation Models work on iOS 26+ today.
+
 ### Unsupported
 
 | Platform | Fallback Behavior |
@@ -26,7 +37,8 @@ On-device AI for Expo apps. Run language models locally—no API keys, no cloud,
 - **Privacy-first** — All inference happens on-device; no data leaves the user's device
 - **Zero latency** — No network round-trips required
 - **Free forever** — No API costs, rate limits, or subscriptions
-- **Native performance** — Built on Apple Foundation Models (iOS) and Google ML Kit Prompt API (Android)
+- **Gemma 4 on-device** — Download and run Gemma 4 E2B/E4B models directly on Android with GPU acceleration
+- **Native performance** — Built on Apple Foundation Models (iOS), ML Kit (Android), and LiteRT-LM (Gemma 4)
 - **Multi-turn conversations** — Full conversation context support
 - **Streaming support** — Progressive token streaming for responsive UIs
 - **Simple API** — Core functions plus prompt helpers for common tasks

package/android/build.gradle

@@ -43,6 +43,6 @@ android {
 }
 
 dependencies {
-  implementation "com.google.mlkit:genai-prompt:1.0.0-alpha1"
-  implementation "com.google.mediapipe:tasks-genai:0.10.24"
+  implementation "com.google.mlkit:genai-prompt:1.0.0-beta2"
+  implementation "com.google.ai.edge.litertlm:litertlm-android:+"
 }

package/android/src/main/java/expo/modules/aikit/GemmaInferenceClient.kt

@@ -1,8 +1,10 @@
 package expo.modules.aikit
 
 import android.content.Context
-import com.google.mediapipe.tasks.genai.llminference.LlmInference
-import kotlinx.coroutines.CompletableDeferred
+import com.google.ai.edge.litertlm.Engine
+import com.google.ai.edge.litertlm.EngineConfig
+import com.google.ai.edge.litertlm.Conversation
+import com.google.ai.edge.litertlm.Backend
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.sync.Mutex
 import kotlinx.coroutines.sync.withLock
@@ -16,7 +18,7 @@ import java.net.URL
 import java.security.MessageDigest
 
 /**
- * Wrapper around MediaPipe LlmInference for Gemma 4 models.
+ * Wrapper around LiteRT-LM Engine for Gemma 4 models.
  *
  * Concurrency model:
  * - A Mutex guards all state transitions (load, unload, inference).
@@ -28,7 +30,8 @@ import java.security.MessageDigest
 class GemmaInferenceClient(private val context: Context) {
 
   private val mutex = Mutex()
-  private var llmInference: LlmInference? = null
+  private var engine: Engine? = null
+  private var conversation: Conversation? = null
   private var loadedModelId: String? = null
 
   @Volatile
@@ -39,33 +42,50 @@ class GemmaInferenceClient(private val context: Context) {
   // -------------------------------------------------------------------------
 
   /**
-   * Load a model into memory. Unloads any previously loaded model first.
+   * Load a model into memory using LiteRT-LM Engine.
+   * Unloads any previously loaded model first.
    * Caller is responsible for emitting onModelStateChange events.
    */
   suspend fun loadModel(modelId: String, modelPath: String) = mutex.withLock {
     // Unload previous model if different
     if (loadedModelId != null && loadedModelId != modelId) {
-      llmInference?.close()
-      llmInference = null
+      conversation?.close()
+      engine?.close()
+      conversation = null
+      engine = null
       loadedModelId = null
     }
 
-    if (loadedModelId == modelId && llmInference != null) {
+    if (loadedModelId == modelId && engine != null) {
       return@withLock // Already loaded
     }
 
     try {
-      val options = LlmInference.LlmInferenceOptions.builder()
-        .setModelPath(modelPath)
-        .build()
-      llmInference = LlmInference.createFromOptions(context, options)
-      loadedModelId = modelId
+      withContext(Dispatchers.IO) {
+        val engineConfig = EngineConfig(
+          modelPath = modelPath,
+          backend = Backend.GPU()
+        )
+        val newEngine = Engine(engineConfig)
+        newEngine.initialize()
+        val newConversation = newEngine.createConversation()
+
+        engine = newEngine
+        conversation = newConversation
+        loadedModelId = modelId
+      }
     } catch (e: OutOfMemoryError) {
-      llmInference = null
+      conversation?.close()
+      engine?.close()
+      conversation = null
+      engine = null
       loadedModelId = null
       throw RuntimeException("INFERENCE_OOM:$modelId:Device does not have enough memory to load model")
     } catch (e: Exception) {
-      llmInference = null
+      conversation?.close()
+      engine?.close()
+      conversation = null
+      engine = null
       loadedModelId = null
       throw RuntimeException("MODEL_LOAD_FAILED:$modelId:${e.message}")
     }
@@ -75,14 +95,16 @@ class GemmaInferenceClient(private val context: Context) {
    * Unload the current model from memory.
    */
   suspend fun unloadModel() = mutex.withLock {
-    llmInference?.close()
-    llmInference = null
+    conversation?.close()
+    engine?.close()
+    conversation = null
+    engine = null
     loadedModelId = null
   }
 
   fun getLoadedModelId(): String? = loadedModelId
 
-  fun isModelLoaded(): Boolean = llmInference != null
+  fun isModelLoaded(): Boolean = engine != null
 
   // -------------------------------------------------------------------------
   // Inference
@@ -93,14 +115,14 @@ class GemmaInferenceClient(private val context: Context) {
    * The mutex ensures this cannot run concurrently with load/unload.
    */
   suspend fun generateText(prompt: String, systemPrompt: String): String = mutex.withLock {
-    val inference = llmInference
+    val conv = conversation
       ?: throw RuntimeException("MODEL_NOT_DOWNLOADED:${loadedModelId ?: "unknown"}:No model loaded")
 
     val fullPrompt = buildFullPrompt(prompt, systemPrompt)
 
     try {
       withContext(Dispatchers.IO) {
-        inference.generateResponse(fullPrompt)
+        conv.sendMessage(contents = fullPrompt).toString()
       }
     } catch (e: OutOfMemoryError) {
       throw RuntimeException("INFERENCE_OOM:${loadedModelId ?: "unknown"}:Out of memory during inference")
@@ -113,47 +135,37 @@ class GemmaInferenceClient(private val context: Context) {
    * Generate a streaming response. The onChunk callback receives
    * (token=delta, accumulatedText=full, isDone) matching the PromptApiClient contract.
    *
-   * MediaPipe's generateResponseAsync passes accumulated text in its partial result
-   * listener, so we diff against previousText to extract the delta token.
-   *
-   * We use a CompletableDeferred to keep the mutex held until streaming completes,
-   * preventing concurrent load/unload during active inference.
+   * LiteRT-LM's sendMessageAsync() returns a Flow<Message>. Each emission
+   * contains accumulated text, so we diff against previousText to extract
+   * the delta token.
    */
   suspend fun generateTextStream(
     prompt: String,
     systemPrompt: String,
     onChunk: (token: String, accumulatedText: String, isDone: Boolean) -> Unit
   ) = mutex.withLock {
-    val inference = llmInference
+    val conv = conversation
       ?: throw RuntimeException("MODEL_NOT_DOWNLOADED:${loadedModelId ?: "unknown"}:No model loaded")
 
     val fullPrompt = buildFullPrompt(prompt, systemPrompt)
 
     try {
       withContext(Dispatchers.IO) {
-        val completion = CompletableDeferred<String>()
         var previousText = ""
 
-        // MediaPipe streaming: generateResponseAsync calls the listener with
-        // accumulated text (not deltas). We normalize to match PromptApiClient's
-        // (token=delta, accumulatedText=full, isDone) contract.
-        inference.generateResponseAsync(fullPrompt) { partialResult, done ->
-          val accumulated = partialResult ?: ""
+        conv.sendMessageAsync(contents = fullPrompt).collect { message ->
+          val accumulated = message.toString()
           val token = if (accumulated.length > previousText.length) {
             accumulated.substring(previousText.length)
           } else {
             ""
           }
           previousText = accumulated
-          onChunk(token, accumulated, done)
-
-          if (done) {
-            completion.complete(accumulated)
-          }
+          onChunk(token, accumulated, false)
         }
 
-        // Wait until streaming finishes so the mutex stays held
-        completion.await()
+        // Final done event for consistency with PromptApiClient
+        onChunk("", previousText, true)
       }
     } catch (e: OutOfMemoryError) {
       throw RuntimeException("INFERENCE_OOM:${loadedModelId ?: "unknown"}:Out of memory during inference")
@@ -189,8 +201,8 @@ class GemmaInferenceClient(private val context: Context) {
         val modelsDir = File(context.filesDir, "models")
         modelsDir.mkdirs()
 
-        val targetFile = File(modelsDir, "$modelId.gguf")
-        val tempFile = File(modelsDir, "$modelId.gguf.tmp")
+        val targetFile = File(modelsDir, "$modelId.litertlm")
+        val tempFile = File(modelsDir, "$modelId.litertlm.tmp")
 
         try {
           val connection = URL(url).openConnection() as HttpURLConnection
@@ -256,17 +268,19 @@ class GemmaInferenceClient(private val context: Context) {
   suspend fun deleteModelFile(modelId: String) = mutex.withLock {
     // Unload if this model is currently loaded
     if (loadedModelId == modelId) {
-      llmInference?.close()
-      llmInference = null
+      conversation?.close()
+      engine?.close()
+      conversation = null
+      engine = null
       loadedModelId = null
     }
 
-    val modelFile = File(context.filesDir, "models/$modelId.gguf")
+    val modelFile = File(context.filesDir, "models/$modelId.litertlm")
     if (modelFile.exists()) {
       modelFile.delete()
     }
     // Also clean up any partial downloads
-    val tempFile = File(context.filesDir, "models/$modelId.gguf.tmp")
+    val tempFile = File(context.filesDir, "models/$modelId.litertlm.tmp")
     if (tempFile.exists()) {
       tempFile.delete()
     }
@@ -276,14 +290,14 @@ class GemmaInferenceClient(private val context: Context) {
    * Check if a model file exists on disk.
    */
   fun isModelFileDownloaded(modelId: String): Boolean {
-    return File(context.filesDir, "models/$modelId.gguf").exists()
+    return File(context.filesDir, "models/$modelId.litertlm").exists()
   }
 
   /**
    * Get the file path for a downloaded model.
    */
   fun getModelFilePath(modelId: String): String {
-    return File(context.filesDir, "models/$modelId.gguf").absolutePath
+    return File(context.filesDir, "models/$modelId.litertlm").absolutePath
   }
 
   // -------------------------------------------------------------------------

package/build/models.d.ts

@@ -14,7 +14,7 @@ export type ModelRegistryEntry = {
     parameterCount: string;
     /** Quantization variant */
     quantization: string;
-    /** URL to download the GGUF model file */
+    /** URL to download the LiteRT-LM model file */
     downloadUrl: string;
     /** SHA256 hash for integrity verification after download */
     sha256: string;
@@ -23,10 +23,9 @@ export type ModelRegistryEntry = {
     /**
      * Practical context window (max tokens) for this model on constrained devices.
      *
-     * These are conservative defaults, NOT the base model's theoretical max (128k).
-     * On a memory-constrained mobile device running quantized inference, KV cache
-     * cannot fit the full 128k context. These values should be benchmarked and
-     * adjusted during Phase 2 testing with real devices.
+     * These are conservative defaults, NOT the base model's theoretical max.
+     * These values should be benchmarked and adjusted during testing with
+     * real devices.
      */
     contextWindow: number;
     /** Minimum device RAM in bytes required to run this model */

package/build/models.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,MAAM,MAAM,kBAAkB,GAAG;IAC/B,6DAA6D;IAC7D,EAAE,EAAE,MAAM,CAAC;IACX,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,4BAA4B;IAC5B,cAAc,EAAE,MAAM,CAAC;IACvB,2BAA2B;IAC3B,YAAY,EAAE,MAAM,CAAC;IACrB,0CAA0C;IAC1C,WAAW,EAAE,MAAM,CAAC;IACpB,4DAA4D;IAC5D,MAAM,EAAE,MAAM,CAAC;IACf,kCAAkC;IAClC,SAAS,EAAE,MAAM,CAAC;IAClB;;;;;;;OAOG;IACH,aAAa,EAAE,MAAM,CAAC;IACtB,6DAA6D;IAC7D,WAAW,EAAE,MAAM,CAAC;IACpB,sCAAsC;IACtC,kBAAkB,EAAE,CAAC,KAAK,GAAG,SAAS,CAAC,EAAE,CAAC;CAC3C,CAAC;AAEF,eAAO,MAAM,cAAc,EAAE,kBAAkB,EA+B9C,CAAC;AAEF;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,MAAM,GAAG,kBAAkB,GAAG,SAAS,CAEhF"}
+{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,MAAM,MAAM,kBAAkB,GAAG;IAC/B,6DAA6D;IAC7D,EAAE,EAAE,MAAM,CAAC;IACX,gCAAgC;IAChC,IAAI,EAAE,MAAM,CAAC;IACb,4BAA4B;IAC5B,cAAc,EAAE,MAAM,CAAC;IACvB,2BAA2B;IAC3B,YAAY,EAAE,MAAM,CAAC;IACrB,+CAA+C;IAC/C,WAAW,EAAE,MAAM,CAAC;IACpB,4DAA4D;IAC5D,MAAM,EAAE,MAAM,CAAC;IACf,kCAAkC;IAClC,SAAS,EAAE,MAAM,CAAC;IAClB;;;;;;OAMG;IACH,aAAa,EAAE,MAAM,CAAC;IACtB,6DAA6D;IAC7D,WAAW,EAAE,MAAM,CAAC;IACpB,sCAAsC;IACtC,kBAAkB,EAAE,CAAC,KAAK,GAAG,SAAS,CAAC,EAAE,CAAC;CAC3C,CAAC;AAEF,eAAO,MAAM,cAAc,EAAE,kBAAkB,EA+B9C,CAAC;AAEF;;;GAGG;AACH,wBAAgB,gBAAgB,CAAC,OAAO,EAAE,MAAM,GAAG,kBAAkB,GAAG,SAAS,CAEhF"}

package/build/models.js

@@ -10,29 +10,29 @@ export const MODEL_REGISTRY = [
         id: 'gemma-e2b',
         name: 'Gemma 4 E2B',
         parameterCount: '2.3B',
-        quantization: 'Q4_K_M',
-        downloadUrl: 'https://huggingface.co/google/gemma-4-e2b-it-GGUF/resolve/main/gemma-4-e2b-it-Q4_K_M.gguf',
+        quantization: 'mixed-2/4/8-bit',
+        downloadUrl: 'https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm/resolve/main/gemma-4-E2B-it.litertlm',
         sha256: '', // TODO: Fill with actual hash once model file is verified
-        sizeBytes: 1_400_000_000, // ~1.4GB
-        // Conservative limit for 4GB RAM devices. Base model supports 128k but
-        // KV cache won't fit. TODO: Benchmark during Phase 2 testing.
+        sizeBytes: 2_580_000_000, // ~2.58GB
+        // Conservative limit for 4GB RAM devices.
+        // TODO: Benchmark during Phase 2 testing.
         contextWindow: 8_000,
         minRamBytes: 4_000_000_000, // 4GB
-        supportedPlatforms: ['ios', 'android'],
+        supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs
     },
     {
         id: 'gemma-e4b',
         name: 'Gemma 4 E4B',
         parameterCount: '4.5B',
-        quantization: 'Q4_K_M',
-        downloadUrl: 'https://huggingface.co/google/gemma-4-e4b-it-GGUF/resolve/main/gemma-4-e4b-it-Q4_K_M.gguf',
+        quantization: 'mixed-4/8-bit',
+        downloadUrl: 'https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/resolve/main/gemma-4-E4B-it.litertlm',
         sha256: '', // TODO: Fill with actual hash once model file is verified
-        sizeBytes: 2_800_000_000, // ~2.8GB
-        // Conservative limit for 6GB RAM devices. Base model supports 128k but
-        // KV cache won't fit. TODO: Benchmark during Phase 2 testing.
+        sizeBytes: 3_650_000_000, // ~3.65GB
+        // Conservative limit for 6GB RAM devices.
+        // TODO: Benchmark during Phase 2 testing.
         contextWindow: 16_000,
         minRamBytes: 6_000_000_000, // 6GB
-        supportedPlatforms: ['ios', 'android'],
+        supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs
     },
 ];
 /**

package/build/models.js.map

@@ -1 +1 @@
-{"version":3,"file":"models.js","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAgCH,MAAM,CAAC,MAAM,cAAc,GAAyB;IAClD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,QAAQ;QACtB,WAAW,EACT,2FAA2F;QAC7F,MAAM,EAAE,EAAE,EAAE,0DAA0D;QACtE,SAAS,EAAE,aAAa,EAAE,SAAS;QACnC,uEAAuE;QACvE,8DAA8D;QAC9D,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,MAAM;QAClC,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;KACvC;IACD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,QAAQ;QACtB,WAAW,EACT,2FAA2F;QAC7F,MAAM,EAAE,EAAE,EAAE,0DAA0D;QACtE,SAAS,EAAE,aAAa,EAAE,SAAS;QACnC,uEAAuE;QACvE,8DAA8D;QAC9D,aAAa,EAAE,MAAM;QACrB,WAAW,EAAE,aAAa,EAAE,MAAM;QAClC,kBAAkB,EAAE,CAAC,KAAK,EAAE,SAAS,CAAC;KACvC;CACF,CAAC;AAEF;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,OAAe;IAC9C,OAAO,cAAc,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,OAAO,CAAC,CAAC;AACtD,CAAC","sourcesContent":["/**\n * Model Registry\n *\n * Defines all downloadable models known to expo-ai-kit.\n * getDownloadableModels() reads from this registry and enriches\n * each entry with on-device status from the native layer.\n */\n\nexport type ModelRegistryEntry = {\n  /** Unique model identifier used in setModel/downloadModel */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Parameter count label */\n  parameterCount: string;\n  /** Quantization variant */\n  quantization: string;\n  /** URL to download the GGUF model file */\n  downloadUrl: string;\n  /** SHA256 hash for integrity verification after download */\n  sha256: string;\n  /** Download file size in bytes */\n  sizeBytes: number;\n  /**\n   * Practical context window (max tokens) for this model on constrained devices.\n   *\n   * These are conservative defaults, NOT the base model's theoretical max (128k).\n   * On a memory-constrained mobile device running quantized inference, KV cache\n   * cannot fit the full 128k context. These values should be benchmarked and\n   * adjusted during Phase 2 testing with real devices.\n   */\n  contextWindow: number;\n  /** Minimum device RAM in bytes required to run this model */\n  minRamBytes: number;\n  /** Platforms this model can run on */\n  supportedPlatforms: ('ios' | 'android')[];\n};\n\nexport const MODEL_REGISTRY: ModelRegistryEntry[] = [\n  {\n    id: 'gemma-e2b',\n    name: 'Gemma 4 E2B',\n    parameterCount: '2.3B',\n    quantization: 'Q4_K_M',\n    downloadUrl:\n      'https://huggingface.co/google/gemma-4-e2b-it-GGUF/resolve/main/gemma-4-e2b-it-Q4_K_M.gguf',\n    sha256: '', // TODO: Fill with actual hash once model file is verified\n    sizeBytes: 1_400_000_000, // ~1.4GB\n    // Conservative limit for 4GB RAM devices. Base model supports 128k but\n    // KV cache won't fit. TODO: Benchmark during Phase 2 testing.\n    contextWindow: 8_000,\n    minRamBytes: 4_000_000_000, // 4GB\n    supportedPlatforms: ['ios', 'android'],\n  },\n  {\n    id: 'gemma-e4b',\n    name: 'Gemma 4 E4B',\n    parameterCount: '4.5B',\n    quantization: 'Q4_K_M',\n    downloadUrl:\n      'https://huggingface.co/google/gemma-4-e4b-it-GGUF/resolve/main/gemma-4-e4b-it-Q4_K_M.gguf',\n    sha256: '', // TODO: Fill with actual hash once model file is verified\n    sizeBytes: 2_800_000_000, // ~2.8GB\n    // Conservative limit for 6GB RAM devices. Base model supports 128k but\n    // KV cache won't fit. TODO: Benchmark during Phase 2 testing.\n    contextWindow: 16_000,\n    minRamBytes: 6_000_000_000, // 6GB\n    supportedPlatforms: ['ios', 'android'],\n  },\n];\n\n/**\n * Look up a model registry entry by ID.\n * Returns undefined if not found.\n */\nexport function getRegistryEntry(modelId: string): ModelRegistryEntry | undefined {\n  return MODEL_REGISTRY.find((m) => m.id === modelId);\n}\n"]}
+{"version":3,"file":"models.js","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AA+BH,MAAM,CAAC,MAAM,cAAc,GAAyB;IAClD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,iBAAiB;QAC/B,WAAW,EACT,uGAAuG;QACzG,MAAM,EAAE,EAAE,EAAE,0DAA0D;QACtE,SAAS,EAAE,aAAa,EAAE,UAAU;QACpC,0CAA0C;QAC1C,0CAA0C;QAC1C,aAAa,EAAE,KAAK;QACpB,WAAW,EAAE,aAAa,EAAE,MAAM;QAClC,kBAAkB,EAAE,CAAC,SAAS,CAAC,EAAE,uCAAuC;KACzE;IACD;QACE,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,aAAa;QACnB,cAAc,EAAE,MAAM;QACtB,YAAY,EAAE,eAAe;QAC7B,WAAW,EACT,uGAAuG;QACzG,MAAM,EAAE,EAAE,EAAE,0DAA0D;QACtE,SAAS,EAAE,aAAa,EAAE,UAAU;QACpC,0CAA0C;QAC1C,0CAA0C;QAC1C,aAAa,EAAE,MAAM;QACrB,WAAW,EAAE,aAAa,EAAE,MAAM;QAClC,kBAAkB,EAAE,CAAC,SAAS,CAAC,EAAE,uCAAuC;KACzE;CACF,CAAC;AAEF;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,OAAe;IAC9C,OAAO,cAAc,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,OAAO,CAAC,CAAC;AACtD,CAAC","sourcesContent":["/**\n * Model Registry\n *\n * Defines all downloadable models known to expo-ai-kit.\n * getDownloadableModels() reads from this registry and enriches\n * each entry with on-device status from the native layer.\n */\n\nexport type ModelRegistryEntry = {\n  /** Unique model identifier used in setModel/downloadModel */\n  id: string;\n  /** Human-readable model name */\n  name: string;\n  /** Parameter count label */\n  parameterCount: string;\n  /** Quantization variant */\n  quantization: string;\n  /** URL to download the LiteRT-LM model file */\n  downloadUrl: string;\n  /** SHA256 hash for integrity verification after download */\n  sha256: string;\n  /** Download file size in bytes */\n  sizeBytes: number;\n  /**\n   * Practical context window (max tokens) for this model on constrained devices.\n   *\n   * These are conservative defaults, NOT the base model's theoretical max.\n   * These values should be benchmarked and adjusted during testing with\n   * real devices.\n   */\n  contextWindow: number;\n  /** Minimum device RAM in bytes required to run this model */\n  minRamBytes: number;\n  /** Platforms this model can run on */\n  supportedPlatforms: ('ios' | 'android')[];\n};\n\nexport const MODEL_REGISTRY: ModelRegistryEntry[] = [\n  {\n    id: 'gemma-e2b',\n    name: 'Gemma 4 E2B',\n    parameterCount: '2.3B',\n    quantization: 'mixed-2/4/8-bit',\n    downloadUrl:\n      'https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm/resolve/main/gemma-4-E2B-it.litertlm',\n    sha256: '', // TODO: Fill with actual hash once model file is verified\n    sizeBytes: 2_580_000_000, // ~2.58GB\n    // Conservative limit for 4GB RAM devices.\n    // TODO: Benchmark during Phase 2 testing.\n    contextWindow: 8_000,\n    minRamBytes: 4_000_000_000, // 4GB\n    supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs\n  },\n  {\n    id: 'gemma-e4b',\n    name: 'Gemma 4 E4B',\n    parameterCount: '4.5B',\n    quantization: 'mixed-4/8-bit',\n    downloadUrl:\n      'https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/resolve/main/gemma-4-E4B-it.litertlm',\n    sha256: '', // TODO: Fill with actual hash once model file is verified\n    sizeBytes: 3_650_000_000, // ~3.65GB\n    // Conservative limit for 6GB RAM devices.\n    // TODO: Benchmark during Phase 2 testing.\n    contextWindow: 16_000,\n    minRamBytes: 6_000_000_000, // 6GB\n    supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs\n  },\n];\n\n/**\n * Look up a model registry entry by ID.\n * Returns undefined if not found.\n */\nexport function getRegistryEntry(modelId: string): ModelRegistryEntry | undefined {\n  return MODEL_REGISTRY.find((m) => m.id === modelId);\n}\n"]}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "expo-ai-kit",
-  "version": "0.2.1",
-  "description": "Expo AI Kit module",
+  "version": "0.3.1",
+  "description": "On-device AI for Expo apps — run Gemma 4, Apple Foundation Models, and ML Kit locally with zero API keys",
   "main": "build/index.js",
   "types": "build/index.d.ts",
   "files": [
@@ -31,6 +31,14 @@
   "keywords": [
     "react-native",
     "expo",
+    "gemma",
+    "gemma-4",
+    "on-device-ai",
+    "llm",
+    "litert",
+    "apple-foundation-models",
+    "ml-kit",
+    "local-inference",
     "expo-ai-kit",
     "ExpoAiKit"
   ],

package/src/models.ts

@@ -15,7 +15,7 @@ export type ModelRegistryEntry = {
   parameterCount: string;
   /** Quantization variant */
   quantization: string;
-  /** URL to download the GGUF model file */
+  /** URL to download the LiteRT-LM model file */
   downloadUrl: string;
   /** SHA256 hash for integrity verification after download */
   sha256: string;
@@ -24,10 +24,9 @@ export type ModelRegistryEntry = {
   /**
    * Practical context window (max tokens) for this model on constrained devices.
    *
-   * These are conservative defaults, NOT the base model's theoretical max (128k).
-   * On a memory-constrained mobile device running quantized inference, KV cache
-   * cannot fit the full 128k context. These values should be benchmarked and
-   * adjusted during Phase 2 testing with real devices.
+   * These are conservative defaults, NOT the base model's theoretical max.
+   * These values should be benchmarked and adjusted during testing with
+   * real devices.
    */
   contextWindow: number;
   /** Minimum device RAM in bytes required to run this model */
@@ -41,31 +40,31 @@ export const MODEL_REGISTRY: ModelRegistryEntry[] = [
     id: 'gemma-e2b',
     name: 'Gemma 4 E2B',
     parameterCount: '2.3B',
-    quantization: 'Q4_K_M',
+    quantization: 'mixed-2/4/8-bit',
     downloadUrl:
-      'https://huggingface.co/google/gemma-4-e2b-it-GGUF/resolve/main/gemma-4-e2b-it-Q4_K_M.gguf',
+      'https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm/resolve/main/gemma-4-E2B-it.litertlm',
     sha256: '', // TODO: Fill with actual hash once model file is verified
-    sizeBytes: 1_400_000_000, // ~1.4GB
-    // Conservative limit for 4GB RAM devices. Base model supports 128k but
-    // KV cache won't fit. TODO: Benchmark during Phase 2 testing.
+    sizeBytes: 2_580_000_000, // ~2.58GB
+    // Conservative limit for 4GB RAM devices.
+    // TODO: Benchmark during Phase 2 testing.
     contextWindow: 8_000,
     minRamBytes: 4_000_000_000, // 4GB
-    supportedPlatforms: ['ios', 'android'],
+    supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs
   },
   {
     id: 'gemma-e4b',
     name: 'Gemma 4 E4B',
     parameterCount: '4.5B',
-    quantization: 'Q4_K_M',
+    quantization: 'mixed-4/8-bit',
     downloadUrl:
-      'https://huggingface.co/google/gemma-4-e4b-it-GGUF/resolve/main/gemma-4-e4b-it-Q4_K_M.gguf',
+      'https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm/resolve/main/gemma-4-E4B-it.litertlm',
     sha256: '', // TODO: Fill with actual hash once model file is verified
-    sizeBytes: 2_800_000_000, // ~2.8GB
-    // Conservative limit for 6GB RAM devices. Base model supports 128k but
-    // KV cache won't fit. TODO: Benchmark during Phase 2 testing.
+    sizeBytes: 3_650_000_000, // ~3.65GB
+    // Conservative limit for 6GB RAM devices.
+    // TODO: Benchmark during Phase 2 testing.
     contextWindow: 16_000,
     minRamBytes: 6_000_000_000, // 6GB
-    supportedPlatforms: ['ios', 'android'],
+    supportedPlatforms: ['android'], // iOS waiting for LiteRT-LM Swift APIs
   },
 ];