Wondering about Google Multi-Token prediction, why isn't this being implemented into every new major model ? Is the 750 token/s achieved using this technique ?