Meta/Llama

I like how they release multiple sizes, but what if they put all their efforts for data and compute into a single model without distributing it among several? Like pick a reasonable size, 14b (like the new Qwen 2.5) or 22b (Mistral Small 2409), and just throw everything at it. I just want to see how much of a leap it would be for that single model size given meta's absurd amount of H100's.