OnDevice puts language, vision, and speech models on your iPhone

Field notes

The project separates its flagship iOS workbench from a server companion and an alternate Core AI runtime. The workbench includes memory-fit checks before loading models, while the server exposes OpenAI-, Anthropic-, and Ollama-compatible routes behind a bearer key. The page lists MLX, llama.cpp, whisper.cpp, and Core ML, and gives a concrete example of a 2.3 GB 4-bit Qwen3-4B 2507 model.