Developers achieved 44–124 tokens per second by aggressively quantizing weights and distributing layers across system RAM and high-speed NVMe storage.