Also right now Sol 5.6 Max is super slow but if it were way faster on a chip (like Taalas' Llama 8b demo) then it would be an extreme value multiplier. But the model is so large that "baking it onto a chip" doesn't seem straightforward.