[azinman2]

Not to mention latency

[disgruntledphd2]

And grounding for the model. Smaller models with tend to hallucinate a little less (anecdotally).