Briefing · 2026-08-06

技术简报 - 2026-08-06

1 immediate1 worth tracking1 source
On this page6 sections

The briefing's title, the editors' summary and their notes are written in Chinese, as are skill and knowledge entries.

Overview

#

今天只有一条新信号,所以这期把端侧的两半放在一起看。Liquid AI 的 LFM2.5-2.6B 把一个能跑完整工具循环的智能体模型压到 2.6B、不到 2.5GB 内存,手机上约 30 tok/s——但值得记住的不是尺寸,而是它的后训练最后一段直接在 OpenClaw、Hermes Agent 这类真实外壳内部跑强化学习,用一个代理把外壳当黑盒、只抓 token 级轨迹。适配的方向因此反了过来:不是你改流程去迁就模型,而是模型在你已经在跑的流程里学。

配的第二条是一周前的 LFM2.5-Encoders,同一代的另一半:编码器那条服务的是输出为标签的负载(意图路由、安全过滤、分类),今天这条服务的是要调用工具、走多步的负载。两条合起来才是「端侧能干什么」的完整答案,单看任何一条都会偏。

最后留一句不该被跑分盖住的话:文章底下有读者反馈,模型从 MCP 工具取回格式错误的数据时没有报错,而是自信地把错的当作事实讲了出来,下一步直接吃掉这个错误。工具调用格式正确,不等于结果正确——要把这类模型放进真实流程,护栏得建在工具层。

Immediate attention

#

The highest-rated published signals; the editors' pinned items come first.

Worth tracking

#

Worth following, but each needs more context or confirmation before it becomes a priority.

Sources

#

The public sources behind the published signals in this briefing.

Liquid AI

Hugging Face Blog - 2026-08-04

Open source

Subscribe

The public feeds carry published briefings only, never drafts or archived ones. They are in Chinese.