It is playing against a language model, but you're right that it's a separate (Gemini) model that plays the moves.

It has an estimated blitz elo of 2820 (see MAV-small in Table 2 of our paper: storage.googleapis.com/deepmind-med...)

Give it a try at: t.co/R0230hFCdI (requires Gemini Advanced subscription but you can get 1 month for free).

The experiment is powered by the MAV–small model from our new paper: deepmind.google/research/pub...