Local AI Bench

LocalAI Bench is an on-device benchmark app that runs AI models directly on your iPhone and measures their performance and capabilities. All inference happens inside your device — your conversations and personal data never leave it.

■ On-Device Execution

• Download open AI models (Llama, Qwen, Gemma, Phi, and more) and run them right on your device
• Works fully offline once a model is downloaded

■ Real-Time Performance Metrics

• Generation speed (tok/s), time-to-first-token (TTFT), peak memory (RAM), and thermal state — displayed live while you chat
• Automated benchmark with a standard methodology: repeated measurements of text generation, prompt processing, translation, coding, and math speed

■ Capability Evaluation

• Auto-graded tasks: math, coding, multilingual translation (Korean↔English/Japanese/Chinese), summarization, reasoning, and vision
• Compare capability, speed, and memory at a glance (choose 2-axis, composite, or efficiency scoring)

■ Vision & Speech Models

• Vision: attach photos or recognize objects through the live camera
• Speech: compare Whisper (tiny/base/small) transcription speed (RTF)

■ Compare & Manage Models

• A/B head-to-head between two models, stress tests (thermal throttling curve)
• Search models by capability (vision, coding, reasoning) or vendor and add them, with automatic iOS compatibility checks

■ Web Search (optional)

• When enabled, fetches Wikipedia references for the model — ask about recent events
• Inference stays 100% on-device; automatically skipped when offline

Pure on-device AI — no cloud APIs, no fees, no subscriptions. Runnable models depend on your device's performance; larger models are recommended for devices with more memory.