“The RWKV paper demonstrated training of its architecture up to a scale of 14 billion parameters.”