The models vary depending on the task. DeepSeek distilled has been a favorite for the past several months.
I use various smaller (~3B) models for simpler tasks.