feat(models): move the default LLM to Qwen3.5 0.8B

src/components/ModelLoaderWidget.tsx

@@ -11,6 +11,11 @@ import { AppColors } from '../theme';
 interface ModelLoaderWidgetProps {
   title: string;
   subtitle: string;
+  /**
+   * "Qwen3.5 0.8B Q4_K_M · Alibaba". Optional so a caller with no single model
+   * behind it (the voice pipeline loads three) can leave it off.
+   */
+  modelCredit?: string;
   icon: string;
   accentColor: string;
   isDownloading: boolean;
@@ -22,6 +27,7 @@ interface ModelLoaderWidgetProps {
 export const ModelLoaderWidget: React.FC<ModelLoaderWidgetProps> = ({
   title,
   subtitle,
+  modelCredit,
   accentColor,
   isDownloading,
   isLoading,
@@ -46,6 +52,9 @@ export const ModelLoaderWidget: React.FC<ModelLoaderWidgetProps> = ({
 
         <Text style={styles.title}>{title}</Text>
         <Text style={styles.subtitle}>{subtitle}</Text>
+        {modelCredit ? (
+          <Text style={[styles.modelCredit, { color: accentColor }]}>{modelCredit}</Text>
+        ) : null}
 
         {(isDownloading || isLoading) && (
           <View style={styles.loadingContainer}>
@@ -125,9 +134,15 @@ const styles = StyleSheet.create({
     fontSize: 14,
     color: AppColors.textSecondary,
     textAlign: 'center',
-    marginBottom: 32,
+    marginBottom: 8,
     lineHeight: 20,
   },
+  modelCredit: {
+    fontSize: 13,
+    fontWeight: '600',
+    textAlign: 'center',
+    marginBottom: 24,
+  },
   loadingContainer: {
     alignItems: 'center',
     marginVertical: 24,

src/screens/ChatScreen.tsx

@@ -13,7 +13,7 @@ import LinearGradient from 'react-native-linear-gradient';
 import { RunAnywhere } from '@runanywhere/core';
 import type { GenerationEvent, GenerationResult } from '@runanywhere/core';
 import { AppColors } from '../theme';
-import { useModelService } from '../services/ModelService';
+import { useModelService, MODEL_CREDITS } from '../services/ModelService';
 import { ChatMessageBubble, ChatMessage, ModelLoaderWidget } from '../components';
 
 export const ChatScreen: React.FC = () => {
@@ -144,6 +144,7 @@ export const ChatScreen: React.FC = () => {
   if (!modelService.isLLMLoaded) {
     return (
       <ModelLoaderWidget
+        modelCredit={MODEL_CREDITS.llm}
         title="LLM Model Required"
         subtitle="Download and load the language model to start chatting"
         icon="chat"

src/screens/SpeechToTextScreen.tsx

@@ -14,7 +14,7 @@ import LinearGradient from 'react-native-linear-gradient';
 import { RunAnywhere, AudioInputs } from '@runanywhere/core';
 import { ModelCategory } from '@runanywhere/proto-ts/model_types';
 import { AppColors } from '../theme';
-import { useModelService } from '../services/ModelService';
+import { useModelService, MODEL_CREDITS } from '../services/ModelService';
 import { ModelLoaderWidget, AudioVisualizer } from '../components';
 
 // Native Audio Module - records in WAV format (16kHz mono) optimal for Whisper STT
@@ -191,6 +191,7 @@ export const SpeechToTextScreen: React.FC = () => {
   if (!modelService.isSTTLoaded) {
     return (
       <ModelLoaderWidget
+        modelCredit={MODEL_CREDITS.stt}
         title="STT Model Required"
         subtitle="Download and load the speech recognition model"
         icon="mic"

src/screens/TextToSpeechScreen.tsx

@@ -11,7 +11,7 @@ import LinearGradient from 'react-native-linear-gradient';
 import { RunAnywhere } from '@runanywhere/core';
 import type { SpeechHandle } from '@runanywhere/core';
 import { AppColors } from '../theme';
-import { useModelService } from '../services/ModelService';
+import { useModelService, MODEL_CREDITS } from '../services/ModelService';
 import { ModelLoaderWidget } from '../components';
 
 const SAMPLE_TEXTS = [
@@ -74,6 +74,7 @@ export const TextToSpeechScreen: React.FC = () => {
   if (!modelService.isTTSLoaded) {
     return (
       <ModelLoaderWidget
+        modelCredit={MODEL_CREDITS.tts}
         title="TTS Voice Required"
         subtitle="Download and load the voice synthesis model"
         icon="volume"

src/screens/ToolCallingScreen.tsx

@@ -14,7 +14,7 @@ import LinearGradient from 'react-native-linear-gradient';
 import { generateWithTools } from '@runanywhere/core';
 import type { ToolCallingResult } from '@runanywhere/core';
 import { AppColors } from '../theme';
-import { useModelService } from '../services/ModelService';
+import { useModelService, MODEL_CREDITS } from '../services/ModelService';
 import { ModelLoaderWidget } from '../components';
 import { DEMO_TOOLS, registerDemoTools } from '../utils/chatSampleTools';
 
@@ -138,6 +138,7 @@ export const ToolCallingScreen: React.FC = () => {
   if (!modelService.isLLMLoaded) {
     return (
       <ModelLoaderWidget
+        modelCredit={MODEL_CREDITS.llm}
         title="LLM Model Required"
         subtitle="Download and load a language model to test tool calling"
         icon="tools"

src/screens/VisionScreen.tsx

@@ -11,7 +11,7 @@ import {
 } from 'react-native';
 import * as RNFS from 'react-native-fs';
 import { AppColors } from '../theme';
-import { useModelService } from '../services/ModelService';
+import { useModelService, MODEL_CREDITS } from '../services/ModelService';
 import { ModelLoaderWidget } from '../components';
 import { VLMService } from '../services/VLMService';
 
@@ -103,6 +103,7 @@ export const VisionScreen: React.FC = () => {
   if (!modelService.isVLMLoaded) {
     return (
       <ModelLoaderWidget
+        modelCredit={MODEL_CREDITS.vlm}
         title="Vision Model Required"
         subtitle="Download and load the vision-language model to describe images"
         icon="vision"

src/services/ModelService.tsx

@@ -1,5 +1,7 @@
 import React, { createContext, useContext, useState, useCallback } from 'react';
 import { Platform } from 'react-native';
+
+import { modelCredit } from './modelOrg';
 import { RunAnywhere } from '@runanywhere/core';
 import {
   ModelCategory,
@@ -11,12 +13,32 @@ import {
 // Model IDs - matching sample app model registry
 // See: runanywhere-sdks/examples/react-native/RunAnywhereAI/src/services/ModelCatalogBootstrap.ts
 export const MODEL_IDS = {
-  llm: 'lfm2-350m-q8_0', // LiquidAI LFM2 - fast and efficient
+  llm: 'qwen3.5-0.8b-q4_k_m', // Qwen3.5 - smallest current-generation chat model
   vlm: 'smolvlm-500m-instruct-q8_0', // SmolVLM - ultra-light vision model
   stt: 'sherpa-onnx-whisper-tiny.en',
   tts: 'vits-piper-en_US-lessac-medium',
 } as const;
 
+/** Display names, kept beside the ids they belong to. */
+export const MODEL_NAMES = {
+  llm: 'Qwen3.5 0.8B Q4_K_M',
+  vlm: 'SmolVLM 500M Instruct',
+  stt: 'Sherpa Whisper Tiny (ONNX)',
+  tts: 'Piper TTS (US English - Medium)',
+} as const;
+
+/**
+ * "Qwen3.5 0.8B Q4_K_M · Alibaba", for the loader screens. A starter that only
+ * says "the language model" leaves the reader with no idea what is about to be
+ * downloaded or who published it.
+ */
+export const MODEL_CREDITS = {
+  llm: modelCredit(MODEL_IDS.llm, MODEL_NAMES.llm),
+  vlm: modelCredit(MODEL_IDS.vlm, MODEL_NAMES.vlm),
+  stt: modelCredit(MODEL_IDS.stt, MODEL_NAMES.stt),
+  tts: modelCredit(MODEL_IDS.tts, MODEL_NAMES.tts),
+} as const;
+
 /**
  * `ModelInfo.isDownloaded` was deleted from the IDL; `registryStatus` is the
  * single downloaded-ness signal now.
@@ -334,22 +356,22 @@ export const ModelServiceProvider: React.FC<ModelServiceProviderProps> = ({ chil
  * runanywhere-sdks/examples/react-native/RunAnywhereAI/src/services/ModelCatalogBootstrap.ts
  */
 export const registerDefaultModels = async () => {
-  // LLM Model - LiquidAI LFM2 350M (fast, efficient, great for mobile)
+  // LLM Model - Qwen3.5 0.8B, the smallest current-generation chat model.
   await RunAnywhere.models.register({
     id: MODEL_IDS.llm,
-    name: 'LiquidAI LFM2 350M Q8_0',
-    url: 'https://huggingface.co/LiquidAI/LFM2-350M-GGUF/resolve/main/LFM2-350M-Q8_0.gguf',
+    name: 'Qwen3.5 0.8B Q4_K_M',
+    url: 'https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/main/Qwen3.5-0.8B-Q4_K_M.gguf',
     framework: InferenceFramework.INFERENCE_FRAMEWORK_LLAMA_CPP,
-    memoryRequirementBytes: 400_000_000,
+    memoryRequirementBytes: 900_000_000,
   });
 
-  // Also add SmolLM2 as alternative smaller model
+  // A smaller alternative for low-memory devices.
   await RunAnywhere.models.register({
-    id: 'smollm2-360m-q8_0',
-    name: 'SmolLM2 360M Q8_0',
-    url: 'https://huggingface.co/prithivMLmods/SmolLM2-360M-GGUF/resolve/main/SmolLM2-360M.Q8_0.gguf',
+    id: 'lfm2.5-230m-q4_k_m',
+    name: 'LiquidAI LFM2.5 230M Q4_K_M',
+    url: 'https://huggingface.co/LiquidAI/LFM2.5-230M-GGUF/resolve/main/LFM2.5-230M-Q4_K_M.gguf',
     framework: InferenceFramework.INFERENCE_FRAMEWORK_LLAMA_CPP,
-    memoryRequirementBytes: 500_000_000,
+    memoryRequirementBytes: 190_000_000,
   });
 
   // VLM Model - SmolVLM 500M (ultra-lightweight vision-language model, ~600MB)

src/services/modelOrg.ts

@@ -0,0 +1,72 @@
+/**
+ * The publisher behind a model, so the loader can say who made the thing it is
+ * about to download instead of just "the language model".
+ *
+ * The rule table is a hand-kept copy of the same table in iOS
+ * (`ModelOrgCatalog`), Android (`ModelTaxonomy`), Web, Electron and Flutter.
+ * Five copies is a known cost; the alternative is a field on the catalog row,
+ * which is a commons change. Keep them in step when a family is added.
+ */
+
+export interface ModelOrg {
+  /** Stable key, e.g. "nvidia". */
+  key: string;
+  /** Consumer-facing name, e.g. "NVIDIA". */
+  name: string;
+}
+
+/**
+ * Ordered matchers against the lowercased "id + name" haystack. First match
+ * wins, so a specific publisher precedes the family it would otherwise be
+ * swallowed by: NVIDIA before Meta so Nemotron stays NVIDIA, DeepSeek before
+ * Alibaba so the R1 Qwen distills stay DeepSeek.
+ */
+const ORG_MATCHERS: ReadonlyArray<{ key: string; name: string; test: RegExp }> = [
+  {
+    key: 'nvidia',
+    name: 'NVIDIA',
+    test: /nemotron|nemoguard|cosmos|canary|parakeet|nv[_-]embed|nv_rerank|nvidia|sortformer/,
+  },
+  { key: 'deepseek', name: 'DeepSeek', test: /deepseek/ },
+  { key: 'prism', name: 'Prism', test: /bonsai|prismml|prism-?ml/ },
+  { key: 'deepgrove', name: 'Deepgrove', test: /maple/ },
+  { key: 'ibm', name: 'IBM', test: /granite/ },
+  // `fara` rides with Microsoft's `phi`: Fara1.5 ships mirrored under our own HF
+  // org, so the catalog row names no upstream publisher. Filing it by its own
+  // name beats guessing one into a UI label.
+  { key: 'microsoft', name: 'Microsoft', test: /\bphi\b|fara/ },
+  { key: 'google', name: 'Google', test: /gemma|embeddinggemma|siglip/ },
+  // Muse Glimmer is Meta's, per the catalog row's own name.
+  { key: 'meta', name: 'Meta', test: /llama|muse-glimmer|muse_glimmer/ },
+  { key: 'alibaba', name: 'Alibaba', test: /qwen/ },
+  { key: 'liquid', name: 'Liquid AI', test: /lfm2/ },
+  { key: 'mistral', name: 'Mistral AI', test: /mistral|ministral/ },
+  { key: 'hugging-face', name: 'Hugging Face', test: /smollm|smolvlm/ },
+  { key: 'openai', name: 'OpenAI', test: /whisper/ },
+  { key: 'zhipu', name: 'Zhipu AI', test: /\bglm\b|glm-/ },
+  {
+    key: 'open-source',
+    name: 'Open source',
+    test: /internvl|moonshine|melo|kokoro|kitten|piper|vits|silero|vad|minilm|supertonic|segformer/,
+  },
+];
+
+const FALLBACK_ORG: ModelOrg = { key: 'open-source', name: 'Open source' };
+
+/**
+ * The publisher for one model. Never throws; an unrecognised name reads as
+ * community rather than guessing a company.
+ */
+export function modelOrg(id: string, name: string): ModelOrg {
+  const haystack = `${id} ${name}`.toLowerCase();
+  const match = ORG_MATCHERS.find((org) => org.test.test(haystack));
+  return match ?? FALLBACK_ORG;
+}
+
+/**
+ * "Qwen3.5 0.8B Q4_K_M · Alibaba" — what the loader shows so a reader knows
+ * what is about to land on their device and who made it.
+ */
+export function modelCredit(id: string, name: string): string {
+  return `${name} · ${modelOrg(id, name).name}`;
+}