they claim it's able to run models with 200B parameters on a single node and 400B when paired with another node