The 1.3b model is amazing for real time code complete, it's fast enough to be a better intellisense.
Another model you should try is magicoder 6.7b ds (based on deepseek coder). After playing with it for a couple weeks, I think it gives slightly better results than the equivalent deepseek model.