#npu

Offline operation, latency, privacy, hardware acceleration, updates and fleet deployment, measured on a laptop: first tokens in 0.1 to 0.2 s, no outside connections, and a model update that refuses a 1-byte corruption.