"Have you personally used any of the latest batch of even smaller local models?"

No I have not, which is why I asked (it wasn't a rhetorical question). Do you have pointers on what the recent improvements are?