Samuel Zeng(@SamuelZengML):35B parameters. One iPhone. No cloud. We trained Edge8-35B, an ultra-sparse MoE with a jointly trained dynamic expert planner, and built an SSD-streaming inference engine around it. In this demo: 44 tok/s, ~1.06 GB peak memory. A truly usable large-model stack for on-device AI. Model, runtime, and paper: open source soon.

Samuel Zeng

@SamuelZengML

Founder, | On-device foundation models. High intelligence, low memory, low power | MIT TR35

参加 October 2012

235 フォロー中 3.7K ファン

Samuel Zeng@SamuelZengML

2026.08.12 15:30

35B parameters. One iPhone. No cloud. We trained Edge8-35B, an ultra-sparse MoE with a jointly trained dynamic expert planner, and built an SSD-streaming inference engine around it. In this demo: 44 tok/s, ~1.06 GB peak memory. A truly usable large-model stack for on-device AI. Model, runtime, and paper: open source soon.