OpenBMB's LES Lets iPhones Run 35B MoE Models in Real Time

OpenBMB open-sourced LES, a lookahead expert scheduling system that runs a 35B-A3B MoE model losslessly on iPhone 18 Pro at 20.5 tokens/s—6x baseline—without skipping experts or altering routing.

2026-10-20 ~ 2026-10-21 · 2 related posts

• ModelBest's LES runs a 35B MoE on-device at 20 tok/s, 6x the baseline — 面壁智能 · 2026-10-20
• 35B MoE at 20.5 tok/s on an iPhone: OpenBMB's Lookahead Expert Scheduling skips no experts — GasSmooth7439 · 2026-10-21