ternii

ternii runs a real AI model entirely on your iPhone. No account, no server, no API key. Nothing you type leaves the device — unless you switch on the optional web tools, which fetch live results for the model to read and give you more powerful answers. OUR OWN MODEL, NOT A SHRUNK COPY Brain 1.0 was trained from scratch and is ternary from the very first token of training: 3.42 billion parameters, 8 experts with 2 active per token. A 4-bit model on a phone is a compressed copy of something bigger. Brain's weights on your phone are the weights it was trained with, so nothing was lost squeezing it on. Measured on held-out text, the 1 GB file on your phone scores within 0.25% of the 7 GB checkpoint we trained. MEASURED ON REAL PHONES - Up to 118 tokens per second on iPhone 17 Pro (typically 108-118), with the first word in about 190 milliseconds - Under 200 MB in use. The weights are memory-mapped rather than loaded, so they cost nothing the operating system charges you for - Runs on older hardware: a 4-billion-parameter model will not run on a 2017 iPhone X. ternii does, at about 16 tokens per second WORKS WITH NO SIGNAL Put the phone in flight mode and it still answers. The model, the expert knowledge and the routing all live on the device. WHAT IT DOES - Ask anything and get an answer on-device - Attach a document or an image and have it read on your phone - Optional connectors you switch on yourself: calendar, contacts, reminders, and your own IMAP mailbox - Optional web search, news, weather and stock lookups, off by default, and the only features that touch the internet - Routines that chain several steps together - Agent mode for multi-step questions, off by default PRIVACY BY CONSTRUCTION Personal data is read on your phone and answered by the on-device model. Mail syncs from your own server using an app password you provide, never an AI cloud. Only the research tools reach the internet, and only once you turn them on. Every connector is an explicit, revocable opt-in. BEFORE YOU DOWNLOAD The app itself is small, but it downloads a model of roughly 1 GB on first run, so use Wi-Fi. A device with at least 3 GB of RAM is recommended. HONEST ABOUT WHAT THIS IS Brain 1.0 is small, fast and private. It is not as accurate as the very large models that run in a data centre, and it can be wrong. Please check anything that matters.