Interesting. Maybe it's one with thinking always-on and one always-off. Similar to the IBM Granite model with the toggle. And on the Claude website, maybe they have another "small" LLM routing the user prompt / controlling that toggle.

→ View original post on X — @rasbt,