It’s definitely for hardware reasons. They have been aggressively improving the vector math capabilities in their chips, but as anybody who has tried to run a local LLM will tell you, newer hardware works better and you’re always limited in what you can do.