I'm running 3.8 27B locally, and the results from the past few days have been excellent. I find raw speed is less of an issue when you can trust the model more to reach the right result.