Running a 70B model on Apple Silicon, end to end

Local inference has crossed a threshold.

With the right quantization and memory configuration, a 70B-parameter model now runs usefully on a single Apple Silicon machine. This report walks through the full setup — model selection, quantization choices, KV-cache tuning — and reports honest token-throughput numbers rather than cherry-picked peaks.

The goal is reproducibility: every number here should be repeatable on comparable hardware.