“A Kimi Linear model was pretrained with 3 billion activated parameters and 48 billion total parameters.”