Opendev is better, and it's just so much easier convincing your boss to let you use it when all company data doesn't get sent to some American company.

While you notice the difference, it was very surprising to me that for almost all tasks, worse models just means things take longer. Plus, both claude and openai lie when they don't know something. I actually have the impression Qwen coder 30B-A3B is a lot more forthcoming about not knowing something.

But the intelligence of the model doesn't actually make the difference between succeed or fail, and when it's exploring an actual problem, the tokens/second makes more difference than the intelligence. And, well, if it's hard enough, just take a break.

Randall, we'll need an update: https://xkcd.com/303/