面壁智能详解 LES：让 35B MoE 模型在手机上跑出 20+ tokens/s
ModelBest Details LES, the Method Behind a 35B MoE Model Running on Phones at Over 20 Tokens/s
面壁智能把 35B 的 MoE 模型塞进手机，靠 LES 预取专家权重这一招，iPhone 上跑到 20.51 tokens/s，代码还开源了。做端侧部署的人可以直接看实现细节，比单纯量化压缩的路子更值得研究。
面壁智能（ModelBest）公开了 LES（Lookahead Expert Scheduling）方法，这套技术让一个 35B 参数的 MoE 模型在手机端运行时达到 20.51 tokens/s 的生成速度，是其基线的 6.07 倍，相关代码已开源。
LES 的核心机制是从手机闪存（flash）中预取下一层的专家权重：MoE 模型每层只激活少量专家，LES 提前判断并加载即将用到的专家，把闪存读取延迟藏在推理过程中，从而避开端侧设备内存不足的瓶颈。
对端侧 AI 开发者来说，这意味着原本必须压缩到几 B 的小模型才能上的手机，现在可以直接跑 35B 级别的 MoE 模型，代码开源后可直接复现和适配到自己的部署流程中。
ModelBest Details LES, the Method Behind a 35B MoE Model Running on Phones at Over 20 Tokens/s
ModelBest's LES prefetches next-layer experts from phone flash; it reports 20.51 tokens/s on an iPhone, 6.07x its baseline, and has open-sourced the code.
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include ModelBest, MoE, LES, 端侧推理.
Primary source verification: published directly by Pandaily with first-party authority.
