MBP M4 Max 64MB - haven't measured the tokens/sec, feels slower than Claude, but not unbearably
It's not yet perfect, my sense is just that it's near the tipping point where models are efficient enough that running a local model is truly viable