I have two of the M3s due testing of models at work and with exo I can run decent quantization with 1 millon tokens for memory and derailment tests. Slow? Yes,

I have two of the M3s due testing of models at work and with exo I can run decent quantization with 1 millon tokens for memory and derailment tests. Slow? Yes, but ... private. Unconditionally. And recently with https://github.com/antirez/ds4 https://github.com/antirez/ds4 one can use just one system to a very, very decent speed and ttft for chat inference. Again, private.