Better models already exist, this is just proving you can dramatically increase inference speeds / reduce inference costs.

It isn't about model capability - it's about inference hardware. Same smarts, faster.