NobodyWho

NobodyWho is an open-source inference engine for running language, vision and speech models entirely on the device that is using them. Built on llama.cpp, it takes any GGUF model from the Hugging Face Hub and runs it locally — no API keys, no cloud round-trip, and no per-token bill.

What sets it apart from most local-inference libraries is the breadth of its first-class bindings. The same engine is exposed idiomatically to Swift, Kotlin, Flutter/Dart, React Native, Python and the Godot game engine, and the API is deliberately small: load a model from a path or an hf:// reference, call ask(), await the completion. Speech is covered by the same surface — a Stt class transcribes audio through Whisper models, and a Tts class synthesises speech from Kokoro-82M with selectable voices and languages.

Beyond plain chat, NobodyWho ships type-safe tool calling with grammar generated automatically from your function signatures, multimodal (vision) input, voice activity detection, and GPU acceleration through Vulkan and Metal. Packages are published to pub.dev, PyPI, npm and the Godot asset library.

The project's own apps demonstrate how far down the hardware ladder it reaches: NobodyWho Chat runs Gemma, Qwen and LFM models on a phone, NobodyWho Wrist claims to be the first app running AI entirely on an Apple Watch, and NobodyWho Eyes targets Apple Vision Pro.

It is developed by a Copenhagen team that raised a €2M pre-seed from PSV Tech, The Footprint Firm and Norrsken Evolve to pursue local, energy-efficient AI as an alternative to hosted frontier models. The source is free and released under the European Union Public Licence v1.2 — a copyleft licence, so check its terms against your distribution plans before shipping a closed product on top of it.