Llama Compose vs GenBench

Llama Compose is the showcase application for Colombia AI Week, created to highlight the power of on-device AI running natively on Apple Silicon (M1 and later) with Metal acceleration. Through an interactive chat interface, it demonstrates how advanced language models can operate locally on iPhone and Mac with real-time performance. Key features: - Multiple chat modes (Simple & Agent) - On-device AI inference with llama.cpp optimized for Metal - Support for Meta's Llama and Google's Gemma model families - On-device agent with tool calling via Koog.ai - Model download and management - Runs on iOS and Apple Silicon Macs - Real-time, interactive conversation interface Important Disclaimer: This app includes experimental AI functionality. AI-generated responses may contain offensive, inaccurate, or inappropriate content. Users should exercise caution and not rely on this app for critical decisions or sensitive information. The app is provided for educational and demonstration purposes only.

GenBench lets you download, run, and benchmark open-source AI models entirely on your device — no servers, no subscriptions, no data leaving your phone. WHAT YOU CAN DO • Download models directly from Hugging Face in one tap • Chat with AI models completely offline and private • Benchmark speed: tokens per second, first-token latency, memory usage • Compare performance across different models • Submit scores to the global leaderboard and see how your device ranks • Share results with friends WHY GENBENCH Your iPhone has a powerful Neural Engine and GPU — GenBench shows you exactly what it can do. Whether you're a developer choosing a model for your app, a researcher testing quantizations, or just curious how fast AI runs on your hardware, GenBench gives you real numbers in seconds. PRIVACY FIRST Everything runs locally. Your prompts never leave your device. No account required. Leaderboard submissions are anonymous — only your score, model name, and device type are shared. SUPPORTED MODELS Works with GGUF-format models including Llama, Mistral, Phi, Gemma, Qwen, and more. Supports text and vision models.