I have to spend some more time with that to understand where it fails to deliver better but I share the feeling of a lack of impressive results thus far. 

I had a short test and I wasn't impressed at all.

It's still better to run your own, but if you don't have acceptable hardware for it, this could at least help in the meantime. My initial questioning of it didn't feel like it was close to the performance I see on some of the (even smaller) latest generation open source LLMs.