This GPT-4o model seems to generate better responses with fewer computational resources, possibly due to the efficiency of its tokenizer.
I haven’t tested its responses in non-English languages enough yet.
At the very least, it’s certain that the speed of response generation doesn’t mean the generated text is inferior in quality.