4 ms·

There are 4 different representations it uses to maximize the available memory for the conversational context.