Compact 1.1B parameter model trained on 3 trillion tokens achieving impressive performance for its size.