Equal to the 2x3090? Yeah it’s about equal in every way, context windows included.
As for useful at that scale?
I use mine for coding a fair bit, and I don’t find it a detractor overall. It enforces proper API discipline, modularity, and hierarchal abstraction. Perhaps the field of application makes that more important though. (Writing firmware and hardware drivers).
It also brings the advantage of focusing exclusively on the problems that are presented in the limited context, and not wandering off on side quests that it makes up.
I find it works well up to about 1KLOC at a time.
I wouldn’t imply they were equal to commercial models, but I would definitely say that local models are very useful tools.
They are also stable, which is not something I can say for SOTA models. You cal learn how to get the best results from a model and the ground doesn’t move underneath you just when you’re on a roll.
As for useful at that scale?
I use mine for coding a fair bit, and I don’t find it a detractor overall. It enforces proper API discipline, modularity, and hierarchal abstraction. Perhaps the field of application makes that more important though. (Writing firmware and hardware drivers).
It also brings the advantage of focusing exclusively on the problems that are presented in the limited context, and not wandering off on side quests that it makes up.
I find it works well up to about 1KLOC at a time.
I wouldn’t imply they were equal to commercial models, but I would definitely say that local models are very useful tools.
They are also stable, which is not something I can say for SOTA models. You cal learn how to get the best results from a model and the ground doesn’t move underneath you just when you’re on a roll.