[Havoc]

>But from what I've heard it's still at least 15% behind GPT-4 for code

More. But you're comparing the performance of something that runs on a macbook vs something out of a datacenter full of H100s.