The model supports a one-million-token context window and introduces a new sparse attention architecture for coding and ...