Some limited intelligence but strictly focused on the environment and goal. Similarly, local 27B models often outperform large 816B (42B active) models in different tasks.