A local model generating 20 tokens/sec today could potentially reach 40 tokens/sec in many scenarios.