Consitancy response from Model's with Ollama ?

[ilanb]

Hi,

I have problems with the responses of the models used with Ollama, despite the fact that they are 70b models, Tried :

llama3.3:70b-instruct-q4_K_S
nemotron:70b-instruct-q4_K_S
hengwen/Sky-T1-32B-Preview:Q8_0
phi4:14b-fp16

The responses are really not terrible... yet used on other applications, I have good results or even as good as a GPT series.
The question is where could this come from? I'm leaning towards Instructor?

Thanks for advice
Best,