Do distilled reasoning models retain the same quality?