工具与项目 3.0 · 值得看 2026-06-14 · 论文

On-Device LLM Deployment Edge Mobile 2026-06-14

中文翻译 量化谱系已基本完整:从 8-bit 到 1.58-bit 都有"工程可用"方案Sub-billion 架构已被重新定义:deep-and-thin + MoE 成为新范式Mobile NPU 进入"软件/硬件协同"时代 English Original Quantization spectrum is basically complete: from 8-bit to 1.58-bit there are "engineering-ready" solutions.

打开原文回到归档

中文翻译

量化谱系已基本完整:从 8-bit 到 1.58-bit 都有"工程可用"方案Sub-billion 架构已被重新定义:deep-and-thin + MoE 成为新范式Mobile NPU 进入"软件/硬件协同"时代

English Original

Quantization spectrum is basically complete: from 8-bit to 1.58-bit there are "engineering-ready" solutions. Sub-billion architecture has been redefined: deep-and-thin + MoE has become a new paradigm.