Just did a quick test in the https://model.box playground, and it looks like the completion length is noticeably shorter than other models (e.g., gpt-4o). However, the response speed meets expectations..