10-19-2026, 04:30 AM
That matches what I was seeing. I ended up in about the same place using MLX. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
measure twice, benchmark three times