which is basically admission that they will API it, not release the models... again.

The original version on github can only compute the specific systems in the paper.

https://github.com/deepmind/deepmind-research/tree/master/al...

I don't know why scientific publications keep doing PR work for them.