npm install @runanywhere/core @runanywhere/llamacpp
# or
yarn add @runanywhere/core @runanywhere/llamacpp

cd ios && pod install && cd ..

import { RunAnywhere, SDKEnvironment, ModelCategory } from '@runanywhere/core';
import { LlamaCPP } from '@runanywhere/llamacpp';

// 1. Initialize SDK
await RunAnywhere.initialize({
  environment: SDKEnvironment.Development,
});

// 2. Register LlamaCPP backend
LlamaCPP.register();

// 3. Add a model
await LlamaCPP.addModel({
  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',
  memoryRequirement: 500_000_000,
});

// 4. Download model
await RunAnywhere.downloadModel('smollm2-360m-q8_0', (progress) => {
  console.log(`Downloading: ${(progress.progress * 100).toFixed(1)}%`);
});

// 5. Load model
const modelInfo = await RunAnywhere.getModelInfo('smollm2-360m-q8_0');
await RunAnywhere.loadModel(modelInfo.localPath);

// 6. Generate text
const response = await RunAnywhere.chat('What is the capital of France?');
console.log(response);

import { LlamaCPP } from '@runanywhere/llamacpp';

LlamaCPP.register(): void

await RunAnywhere.initialize({ ... });
LlamaCPP.register();  // Now LLM features are available

await LlamaCPP.addModel(options: LlamaCPPModelOptions): Promise<ModelInfo>

interface LlamaCPPModelOptions {
  /**
   * Unique model ID.
   * If not provided, generated from the URL filename.
   */
  id?: string;

  /** Display name for the model */
  name: string;

  /** Download URL for the model (GGUF format) */
  url: string;

  /**
   * Model category.
   * Default: ModelCategory.Language
   */
  modality?: ModelCategory;

  /**
   * Memory requirement in bytes.
   * Used for device capability checks.
   */
  memoryRequirement?: number;

  /**
   * Whether model supports reasoning/thinking tokens.
   * If true, thinking content is extracted from responses.
   */
  supportsThinking?: boolean;
}

// Basic model
await LlamaCPP.addModel({
  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',
  memoryRequirement: 500_000_000,
});

// Larger model
await LlamaCPP.addModel({
  id: 'llama-2-7b-chat-q4_k_m',
  name: 'Llama 2 7B Chat Q4_K_M',
  url: 'https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/resolve/main/llama-2-7b-chat.Q4_K_M.gguf',
  memoryRequirement: 4_000_000_000,
});

// Model with thinking support (e.g., DeepSeek-R1)
await LlamaCPP.addModel({
  id: 'deepseek-r1-distill-qwen-1.5b',
  name: 'DeepSeek R1 Distill Qwen 1.5B',
  url: 'https://huggingface.co/.../deepseek-r1-distill-qwen-1.5b-q8_0.gguf',
  memoryRequirement: 2_000_000_000,
  supportsThinking: true,
});

LlamaCPP.moduleId        // 'llamacpp'
LlamaCPP.moduleName      // 'LlamaCPP'
LlamaCPP.inferenceFramework  // LLMFramework.LlamaCpp
LlamaCPP.capabilities    // ['llm']
LlamaCPP.defaultPriority // 100

const response = await RunAnywhere.chat('Hello!');
console.log(response);

const result = await RunAnywhere.generate(
  'Explain machine learning in simple terms',
  {
    maxTokens: 256,
    temperature: 0.7,
    topP: 0.95,
    systemPrompt: 'You are a helpful teacher.',
    stopSequences: ['\n\n'],
  }
);

console.log('Response:', result.text);
console.log('Tokens:', result.tokensUsed);
console.log('Speed:', result.performanceMetrics.tokensPerSecond, 'tok/s');
console.log('TTFT:', result.performanceMetrics.timeToFirstTokenMs, 'ms');

const streamResult = await RunAnywhere.generateStream(
  'Write a story about a robot',
  { maxTokens: 500 }
);

// Display tokens as they're generated
for await (const token of streamResult.stream) {
  process.stdout.write(token);
}

// Get final metrics
const result = await streamResult.result;
console.log('\nSpeed:', result.performanceMetrics.tokensPerSecond, 'tok/s');

import { RunAnywhere } from '@runanywhere/core';
import { LlamaCPP } from '@runanywhere/llamacpp';

// Register a tool
RunAnywhere.registerTool(
  {
    name: 'calculate',
    description: 'Perform a math calculation',
    parameters: [
      { name: 'expression', type: 'string', description: 'Math expression', required: true },
    ],
  },
  async (args) => {
    const result = eval(args.expression as string); // simplified example
    return { result };
  }
);

// Generate with tools
const result = await RunAnywhere.generateWithTools(
  'What is 42 * 17?',
  {
    autoExecute: true,
    maxToolCalls: 3,
    temperature: 0.7,
    format: 'default', // 'default' for most models, 'lfm2' for Liquid AI models
  }
);
console.log(result.text); // "42 * 17 = 714"

// Load model
await RunAnywhere.loadModel('/path/to/model.gguf');

// Check if loaded
const isLoaded = await RunAnywhere.isModelLoaded();

// Unload to free memory
await RunAnywhere.unloadModel();

// Cancel ongoing generation
await RunAnywhere.cancelGeneration();

packages/llamacpp/
├── src/
│   ├── index.ts                # Package exports
│   ├── LlamaCPP.ts             # Module API (register, addModel)
│   ├── LlamaCppProvider.ts     # Service provider
│   ├── native/
│   │   └── NativeRunAnywhereLlama.ts
│   └── specs/
│       └── RunAnywhereLlama.nitro.ts
├── cpp/
│   ├── HybridRunAnywhereLlama.cpp
│   ├── HybridRunAnywhereLlama.hpp
│   └── bridges/
├── ios/
│   ├── RunAnywhereLlama.podspec
│   └── Frameworks/
│       └── RABackendLLAMACPP.xcframework
├── android/
│   ├── build.gradle
│   └── src/main/jniLibs/
│       └── arm64-v8a/
│           └── librunanywhere_llamacpp.so
└── nitrogen/
    └── generated/