tensai

Theme

Language

Your phone, your AI.

tensai runs open models directly on your Android phone. Chat, reasoning and an agent that can act, with nothing leaving the device unless you ask for it.

Download

Pre-1.0, expect changes

How it works

llama.kt is open too

A tensai chat: a short reply to a question about phone battery, with its speed stats under the replyA tensai chat: a short reply to a question about phone battery, with its speed stats under the reply

On the device

Models run on your phone's CPU or GPU through llama.cpp. Once a model is downloaded, tensai works offline.

Any open model

Import a GGUF file or pick one from Hugging Face. A fit check rates it against your phone's memory.

  • Runs well
  • Tight fit
  • Won't fit

An agent, with guardrails

Tools for dates, maths, memory, timers, calendar and messages. Anything that sends or deletes asks you first. Internet tools stay off until you enable them.

What it looks like

  • The model picker: the loaded model with a check, another model marked Tight fit, then Manage models and Discover modelsThe model picker: the loaded model with a check, another model marked Tight fit, then Manage models and Discover models
    Pick a model. The fit check shows what your phone can run.
  • A chat where the model has reasoned about whether 391 is a prime number: the Reasoning block is open, with products like 17*23 checked step by step, and the answer belowA chat where the model has reasoned about whether 391 is a prime number: the Reasoning block is open, with products like 17*23 checked step by step, and the answer below
    Read the model's reasoning, or fold it away.
  • A tool call waiting for confirmation: tensai wants to forget the fact “my favourite tea is genmaicha”, with Deny and AllowA tool call waiting for confirmation: tensai wants to forget the fact “my favourite tea is genmaicha”, with Deny and Allow
    Anything that deletes or sends waits for your answer.
  • The context sheet: how much of the window is used, the runtime and the agent's tool callsThe context sheet: how much of the window is used, the runtime and the agent's tool calls
    See what fills the context, agent activity included.

Measured

Your phone will differ. Big models are slow on phones; tensai tells you before you download.

Dimensity 900 · Qwen3 1.7B Q4_K_M · CPU

first message once ready~1 s
reply speed~5.5 tok/s
model load~7 s

The tool list is prepared in the background after loading, about a minute on this phone.

Coming next

  • MCP and skills
  • 1.0: stable release
  • API server on your network

Apache 2.0. The app and the engine, llama.kt, are on GitHub.

llama.kt on GitHubRead the docs