Tara Core

Tara Core runs large language models directly on your phone. Nothing you type is sent anywhere. There is no account, no server, and no internet connection needed once a model is downloaded. WHAT MAKES IT DIFFERENT Most on-device AI apps bundle their own engine and their own copy of the model. Three such apps means three engines and three multi-gigabyte downloads. Tara Core is installed once and shared: any app on your phone can use it, so they each ship nothing. WHAT YOU CAN DO • Chat with a model straight away, offline • Choose from a catalogue of open models: Gemma, Qwen, Llama, Phi, SmolLM • See exactly how much memory each model needs before you download it • Watch speed and memory use on the dashboard, or add the home screen widget • Let other apps on your phone use the same engine FOR DEVELOPERS Add local AI to your app in three lines, without shipping an engine or a model: val client = TaraCoreClient(context).apply { connect() } client.chatStream(listOf(ChatMessageParcel("user", "Hello"))) .collect { piece -> print(piece) } Or point any OpenAI-compatible client at http://127.0.0.1:8080/v1 — the standard Python openai package works unmodified. Constrained output is supported, so a small model can be made to answer with exactly one of your options, or with JSON in a shape you specify. PRIVACY Your conversations are never stored and never transmitted. The only network requests the app makes are the model downloads you start, and an optional update check that sends nothing about you. The local server listens on 127.0.0.1 only and is protected by a key generated on your phone. WHAT YOU NEED • A 64-bit phone, which means almost any phone since 2018 • 4 GB of memory for small models; 8 GB is comfortable • Storage for the models you choose, from 100 MB to several GB A NOTE ON SPEED Everything runs on your phone's processor, so larger models are slower. A small model answers at conversation speed; a large one is closer to reading speed. The app tells you what each model needs before you download it. Open source, Apache 2.0: https://github.com/weberq/taracore