OPEN SOURCE OBSERVER
从项目热度走向工程判断：追踪开源工具、模型与开发者生态。
01 / CODE
| 排名 | 项目 | 语言 | 今日新增 | 总 Star |
|---|---|---|---|---|
| 01 | fmtlib/fmtA modern formatting library | C++ | +955 | 25,017 |
| 02 | sponsors/mattpocockSkills for Real Engineers. Straight from my .agents directory. | Shell | +1,576 | 247,183 |
| 03 | NousResearch/hermes-agentThe agent that grows with you | Python | +778 | 240,778 |
| 04 | sponsors/DietrichGebertMakes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote. | JavaScript | +2,138 | 123,279 |
| 05 | anthropics/skillsPublic repository for Agent Skills | Python | +277 | 173,599 |
| 06 | sponsors/affaan-mThe agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond. | JavaScript | +749 | 247,068 |
| 07 | sponsors/JuliusBrussee🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman | Go | +545 | 103,051 |
| 08 | blader/humanizerAgent skill that removes signs of AI-generated writing from text | Python | +1,214 | 41,382 |
| 09 | google-research/timesfmTimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting. | Python | +1,626 | 30,670 |
| 10 | averygan/reclipDownload videos from almost any website. Lightweight, self-hosted media downloader with a clean web UI. | HTML | +123 | 8,319 |
02 / MODELS
| 排名 | 模型 | 任务 | 收藏 | 下载 |
|---|---|---|---|---|
| 01 | Z-Image-Turbo | 文本生成图片 | 2,465 | 791,771 |
| 02 | MiniMax-H3 | — | 1,913 | 125,939 |
| 03 | DeepSeek-R1 | — | 1,379 | 927,253 |
| 04 | Qwen3.8-27B | 视觉多模态理解 | 1,327 | 167,170 |
| 05 | Qwen3-Coder-480B-A35B-Instruct | 文本生成 | 1,092 | 138,489 |
| 06 | Qwen3.8-Flash-Next | 视觉多模态理解 | 1,035 | 23,293 |
| 07 | IndexTTS-2 | 语音合成 | 971 | 203,311 |
| 08 | chatglm3-6b | — | 828 | 1,590,584 |
| 09 | Qwen-7B-Chat | 文本生成 | 757 | 693,072 |
| 10 | Qwen3.8-2.4T-A95B | 文本生成 | 753 | 1,713 |
DEVELOPER VIEW
几个值得追踪的动向：
Agent Skill的“可学性”极高。anthropics/skills官方仓库+第三方生态呈大量结构化小样例，短期可快速学习如何将私有经验工程化注入Agent。
Token效率优化切入点稀缺有效——例如JuliusBrussee的“精简语言”风格已演示最大65%的token削减，低成本赢家。
TimesFM将预测这事开源门禁进一步拉开，适合快速实验；不用微调即可获得短时序预测正确率显著优于统计基准的效果。
Qwen3.8系列在免费开源与多工况可用性之间表现出众，轻量Flash/Nex尺寸尤其适合国内做私有化部署初期评估；humanizer这类去AI味工具则直接叩问内容行业AI产出如何“去痕迹”。
WORTH A STAR
1. anthropics/skills——难得围绕Agent可执行技能（而非模型权重）开源的官方存储库，是Agent生态的工业标准雏形，值得当下持续跟踪、提交并从中找范式灵感。
2. google-research/timesfm——谷歌预训练时序基础模型。不要只看框架，先下手试试它对做短序列任务的可用性；该类方向大概率还会继续吸star。
3. blader/humanizer——宣称能移除AI生成痕迹，一天+1214。它本身是“AI检测对抗”实验的入场券；对编辑、内容创作与出版链路中的人类鉴别标准都值得当作参照物持续观察。
SOURCES & METHOD
GitHub Trending（项目热度、语言、Star） · ModelScope（模型任务、收藏与下载）
抓取日期：2026-09-04。排名是当日快照，指标可能随平台更新而变化。