"The 26 billion parameter model can run on something like the Mac Studio M2 Ultra, which is already a few years old, can still push around 300 tokens per second."