Just released, the latest iteration of tiny LLM by Microsoft, only 4B parameters and outperforming other, bigger models, most notably Mixtral 8x7B. GPT3.5 still quite a bit better, but consider the size difference!
Another clear indicator that we are headed for smaller, local models. I am looking forward what to Apple will be doing here when their new phones come out in fall.
Matthew Berman's review: