A Go AI that runs on your phone — offline

Most Go apps send your moves to a server. TseGo runs the neural network on your device, so the AI works on a plane, in the metro, and with no account.

Why on-device matters

• No latency. The move appears when the engine finishes thinking, not after a round trip.
• Works with no signal. Plane, metro, countryside — the full-strength AI still plays.
• Nothing leaves your phone. Your games against the AI are never sent anywhere.
• No account. Install, open, play.

How it is built

The engine is KataGo, the strongest open-source Go network. Its b18 model is converted to ONNX with fp16 weights and executed through ONNX Runtime, using WebGPU or WASM depending on the device. The model is downloaded once, then cached.

On top of the raw network there is a small time-budgeted search: it expands a real tree with PUCT selection and batched leaf evaluation, so on the strongest setting the AI reads sequences ahead instead of playing a one-move-deep policy guess.

The hard part: making it play badly, believably

Strong engines are easy. A convincing 15 kyu is hard. Cut a strong network's thinking time and it still finds dan-level moves — it just finds them a bit less often. The result feels like a machine that occasionally blunders, not like a human.

TseGo uses KataGo's human-imitation network, conditioned on a target rank. Ask it for 8 kyu and it produces the move an 8 kyu would actually play: the slightly slow shape, the overplay in the corner, the missed tesuji. That makes the strength slider mean something — a "beginner" opponent is genuinely beatable by a beginner.

Strength levels

What still needs a server

Honesty matters here: game analysis — reviewing a finished game move by move at high visit counts — is far heavier than playing, and that part is server-side. Playing, puzzles and the tutorial are fully offline.

See also playing online and the puzzle path.