frontier models are still way better than local models from what I've seen. To get close to them with large context windows and decent performance, you need more than a reasonable machine imo.

I'm hoping local llms start rapidly improving even more though.