3 ms·Model sizes are currently constrained by availability of inference hardware1 points by ipnon 10mo ago