Built with toolnexus-web —
the browser edition of toolnexus.
One import, NexusDecider, is the whole AI side of this game.
GitHub ·
npm
Every move is a Jev-style System One decision from open-jev, a DeBERTa model trained only to decide: the board goes in as one sentence, the moves that do not crash go in as the allowed options, and one forward pass scores them. No text is generated. It runs on your GPU, in this tab. ← back to the demo
Load a model and press “Jev plays”.
The model sees this sentence and only the moves that do not crash — Jev's "allowed options". It never sees the grid, never writes a word, and cannot answer anything else.
Same 60 random boards, measured in this page on WebGPU: a 0.6B chat model reading letter logits moved toward the food 45% of the time — a random safe move manages 49%. This model: 100%, at 36 ms a move.
What the game just ran, with the real state, the moves it offered and what came back.
It updates every move. Everything else on this page is drawing the board.
npm install toolnexus-web — the same call works in your app.