[RNE Rewrite] feat(models)!: resolve the DEFAULT model variant per platform

.cspell-wordlist.txt

@@ -350,3 +350,4 @@ phonemizes
 həlˈoʊ
 NSURL
 backgrounding
+LibriSpeech

apps/computer-vision/app/detection/index.tsx

@@ -16,11 +16,11 @@ import { BoundingBox } from '../../components/BoundingBox';
 const MODEL_OPTIONS: ModelOption[] = [
   {
     label: 'SSDLite 320 MobileNet V3 Large (XNNPACK FP32)',
-    value: models.objectDetection.SSDLITE320_MOBILENET_V3_LARGE.DEFAULT,
+    value: models.objectDetection.SSDLITE320_MOBILENET_V3_LARGE.XNNPACK_FP32,
   },
   {
     label: 'RF-DETR Nano (XNNPACK FP32)',
-    value: models.objectDetection.RFDETR_NANO.DEFAULT,
+    value: models.objectDetection.RFDETR_NANO.XNNPACK_FP32,
   },
   {
     label: 'RF-DETR Nano (CoreML FP16)',

apps/computer-vision/app/keypoint/index.tsx

@@ -16,7 +16,7 @@ import { BoundingBox } from '../../components/BoundingBox';
 const MODEL_OPTIONS: ModelOption[] = [
   {
     label: 'BlazeFace (XNNPACK FP32)',
-    value: models.keypointDetection.BLAZEFACE.DEFAULT,
+    value: models.keypointDetection.BLAZEFACE.XNNPACK_FP32,
   },
   {
     label: 'YOLO26 Pose (XNNPACK FP32)',
@@ -27,8 +27,8 @@ const MODEL_OPTIONS: ModelOption[] = [
     value: models.keypointDetection.RFDETR_KEYPOINT.XNNPACK_FP32,
   },
   {
-    label: 'RF-DETR Keypoint (CoreML FP32)',
-    value: models.keypointDetection.RFDETR_KEYPOINT.COREML_FP32,
+    label: 'RF-DETR Keypoint (CoreML FP16)',
+    value: models.keypointDetection.RFDETR_KEYPOINT.COREML_FP16,
     disabled: Platform.OS !== 'ios',
   },
   {

apps/computer-vision/package.json

@@ -10,8 +10,10 @@
       "instanceSegmentation",
       "keypointDetection",
       "objectDetection",
+      "ocr",
       "semanticSegmentation",
-      "styleTransfer"
+      "styleTransfer",
+      "textToImage"
     ]
   },
   "scripts": {

apps/speech/app/audio-file-transcription/index.tsx

@@ -73,6 +73,31 @@ const MODELS = [
     config: models.speechToText.WHISPER.TINY.MLX_BF16,
     disabled: Platform.OS !== 'ios',
   },
+  {
+    name: 'Tiny Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Tiny Multilingual (Vulkan int8)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_INT8,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Tiny English (Vulkan fp16)',
+    config: models.speechToText.WHISPER.EN.TINY.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Base Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.BASE.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Small Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.SMALL.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
   {
     name: 'Base Multilingual (CPU)',
     config: models.speechToText.WHISPER.BASE.XNNPACK_FP32,

apps/speech/app/microphone-transcription/index.tsx

@@ -66,6 +66,31 @@ const MODELS = [
     config: models.speechToText.WHISPER.TINY.MLX_BF16,
     disabled: Platform.OS !== 'ios',
   },
+  {
+    name: 'Tiny Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Tiny Multilingual (Vulkan int8)',
+    config: models.speechToText.WHISPER.TINY.VULKAN_INT8,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Tiny English (Vulkan fp16)',
+    config: models.speechToText.WHISPER.EN.TINY.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Base Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.BASE.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
+  {
+    name: 'Small Multilingual (Vulkan fp16)',
+    config: models.speechToText.WHISPER.SMALL.VULKAN_FP16,
+    disabled: Platform.OS !== 'android',
+  },
   {
     name: 'Base Multilingual (CPU)',
     config: models.speechToText.WHISPER.BASE.XNNPACK_FP32,

apps/speech/app/text-to-speech/index.tsx

@@ -53,11 +53,14 @@ const STEPS_OPTIONS = [
 const MODEL_OPTIONS = [
   { label: 'XNNPACK (CPU)', value: 'XNNPACK_FP32' as const },
   { label: 'MLX (Apple Silicon)', value: 'MLX_FP32' as const, disabled: Platform.OS !== 'ios' },
+  { label: 'Vulkan (GPU)', value: 'VULKAN_FP16' as const, disabled: Platform.OS !== 'android' },
 ];
 
 function TTSContent() {
   const [text, setText] = useState(SAMPLE_TEXT);
-  const [selectedModel, setSelectedModel] = useState<'XNNPACK_FP32' | 'MLX_FP32'>('XNNPACK_FP32');
+  const [selectedModel, setSelectedModel] = useState<'XNNPACK_FP32' | 'MLX_FP32' | 'VULKAN_FP16'>(
+    Platform.OS === 'android' ? 'VULKAN_FP16' : 'XNNPACK_FP32'
+  );
   const [selectedVoice, setSelectedVoice] = useState<SupertonicDefaultVoiceName>('F1');
   const [selectedLang, setSelectedLang] = useState<speech.SupertonicLanguage>('en');
   const [speed, setSpeed] = useState(1.05);

apps/speech/package.json

@@ -5,6 +5,7 @@
   "react-native-executorch": {
     "features": [
       "vad",
+      "speechToText",
       "textToSpeech"
     ]
   },

docs/docs/01-fundamentals/01-getting-started.md

@@ -213,17 +213,48 @@ The three lists are merged, so you can pair a `features` set with an extra `back
 
 ### Feature → backend / lib mapping
 
-These tasks are available today:
+Each feature provisions the union of the backends its models are published for, so that the `DEFAULT` variant of every model in that family can resolve to the fastest export the device supports (see below).
 
 | Feature | Backends | Extra libs |
 | --- | --- | --- |
+| `llm` | xnnpack, mlx | — |
+| `multimodalLLM` | xnnpack, mlx, vulkan | opencv |
+| `privacyFilter` | xnnpack, mlx | — |
+| `speechToText` | xnnpack, coreml, mlx | — |
+| `textToSpeech` | xnnpack, mlx | phonemis |
+| `vad` | xnnpack | — |
+| `textEmbeddings` | xnnpack, mlx | — |
+| `imageEmbeddings` | xnnpack, coreml, mlx | opencv |
 | `classification` | xnnpack, coreml | opencv |
-| `semanticSegmentation` | xnnpack | opencv |
-| `styleTransfer` | xnnpack, coreml | opencv |
+| `objectDetection` | xnnpack, coreml | opencv |
 | `keypointDetection` | xnnpack, coreml, mlx | opencv |
+| `semanticSegmentation` | xnnpack, coreml | opencv |
+| `instanceSegmentation` | xnnpack, coreml | opencv |
+| `ocr` | xnnpack, coreml, vulkan | opencv |
+| `verticalOCR` | xnnpack | opencv |
+| `styleTransfer` | xnnpack, coreml | opencv |
+| `textToImage` | xnnpack, coreml | opencv |
+| `segmentAnything` | xnnpack, coreml | opencv |
 | `tokenizer` | — | — |
 
-The map also contains forward-looking entries (`llm`, `multimodalLLM`, `speechToText`, `objectDetection`, `ocr`, …) for tasks that are not yet exposed in the JS API; requesting one provisions the right binaries but has no hook to call yet.
+### How the backends you pick change which model runs
+
+Models published for more than one backend expose their exports as named variants next to a `DEFAULT` alias:
+
+```ts
+models.classification.EFFICIENTNET_V2_S.DEFAULT; // resolved for this device
+models.classification.EFFICIENTNET_V2_S.COREML_FP16; // always this export
+models.classification.EFFICIENTNET_V2_S.XNNPACK_INT8;
+```
+
+`DEFAULT` is not a fixed file. It is resolved when the library loads, to the fastest export the device can actually run. A model is only exported to an accelerated backend once it has been shown to run better there, so a published accelerated variant is preferred and XNNPACK is the fallback:
+
+- **iOS device** — Core ML where one exists, otherwise MLX, otherwise XNNPACK.
+- **Android** — Vulkan where one exists, otherwise XNNPACK.
+- **iOS simulator** — XNNPACK. The simulator has no Neural Engine, cannot run Core ML models at all, and MLX ships a device slice only.
+- **All of them, narrowed by your config** — only backends your app downloaded are considered. Trimming `coreml` out of an iOS build moves those `DEFAULT`s to the next best export rather than failing to load.
+
+A handful of models publish both a Core ML and an MLX export. There the two are close enough that the winner is a per-model benchmark result, so the registry pins it explicitly rather than letting the order above decide. Naming any variant directly always overrides the resolution.
 
 ### Platform notes