A collection of Chinese-language tips on using DeepSeek as of Nov 2026: DeepSeek Harness plugins and Creator Mode, free interfaces for connecting to Claude Code, ways to save tokens, and hands-on tests of local deployment and mobile use. Suitable for those who want to use DeepSeek more economically and smoothly.
Not in your language? Paste the link or the text into your AI and let it read it for you.
To save tokens, consider both sides: the paper shows DeepSeek Harness uses far fewer tokens per question than Codex, so before installing a plugin, think through how much context it takes up.
Creator Mode can generate a plugin from one sentence: copy the official prompt and send it. In tests, making one took 10 to 21 minutes and about $0.1 to $0.38.
Prefer the official API: someone tested and found that within the same-priced plan, a larger share went to reasoning. After unsubscribing and switching to the official interface, the value was better.
AMD's China site offers a free DeepSeek interface. Sign up to get a key and use it right away. It is compatible with the Anthropic protocol and can connect to Claude Code.
The barrier to local deployment is falling: the quantized version runs with 64GB of RAM, and FreeToken can run on gaming graphics cards and connect to Claude Code.
You can also use it on your phone: community projects DSH Go and DSH-Android put the engine on Android without root access, allowing local operation or remote control of a computer.
For multitasking, use routing dispatch: the front end only takes requests and replies, and each task opens its own session and working directory to run in parallel without getting mixed up.
Everyday savings: set the free DeepSeek as the main model for Hermes, so it takes over when quota runs out, and it costs almost nothing.
yetone11-02ZHUsing DeepSeek V4.1 Flash in Alma to Find Original High-Definition Wallpapers#1第一轮测试，我就是将原推里的屏幕截图分别给了它们，让它们找到该壁纸的最高清版本，结果如下（图片也按文字的顺序展示各个 Agent 的结果）： 1. Grok bot：用最快的速度找到了推特上的原图，但大小只有 736×736，直接返回给我结束了。DeepSeekImageOriginal post55KLikes 246雨哥向前冲11-02ZHTsinghua-Microsoft Paper Shows DeepSeek Harness Uses Nearly 90% Fewer Tokens#2清华 × 微软亚研这篇新论文，我读完最大的感受是：DSH 被严重低估了！ 论文本来是讲 HarnessMM，一个让 agent 别再"做完一步就从头想一遍"的方法。它拿 Codex 和 DeepSeek Harness两个工业级外壳当实验对象。DeepSeekImageOriginal post40KLikes 337SuSu_酥酥👅11-02ZHDeepSeek Harness Creator Mode Generates a Plugin from One Sentence#3🚨梁圣出招了！DeepSeek Harness 正式上线创作者模式！ 一句话就能把想法变成可直接用的插件，从设计工作台、便签墙、倒计时，到沙盒游戏、3D LLM 可视化、视频剪辑台都能生成。DeepSeekImageOriginal post30KLikes 156SuSu_酥酥👅11-01ZHA List of 8 Worthwhile Plugins for DeepSeek Harness🔥精选！DeepSeek Harness 必装的 8 款插件！让你的 DSH 能力全面升级！ 从找插件到出图、记忆、远程操控、上下文洞察全覆盖，直接把 DSH 拉满： 1️⃣dshfind 插件搜索与发现，按分类找插件，查看项目资料与作者入口。DeepSeekImageOriginal post24KLikes 327SuSu_酥酥👅11-03ZHMeasured Costs of Creating a Plugin with One Sentence in DeepSeek Harness Creation Mode💥牛逼！DeepSeek Harness 创造模式实测：一句话造插件，真的好用！ 刚拿官方提示词实测了一轮，3个插件全跑通： 1️⃣用法极简 复制提示词 → 选「创造模式」→ 发出去。它自己读文档、写代码、打包。DeepSeekImageOriginal post20KLikes 115
DaBin.Binah11-03ZHGet a free key on InferX to call DeepSeek V4.1 Flash白嫖美国公司的DeepSeek，API调用0元🐳 前微软工程师做的AI平台InferX，总部在美国旧金山DeepSeekImageOriginal post20KLikes 199SuSu_酥酥👅11-02ZHWhich Plugins to Install on DeepSeek Harness to Save Tokens and Context🔥务必码住！DeepSeek Harness 这6个插件，直接把Token和上下文从“烧钱”变“可控”！ 四个在抢上下文，三个补能力，三个管观察，装之前先想清楚它替你干什么。DeepSeekImageOriginal post19KLikes 91Cheery11-01ZHDeepSeek Super Agent Uses Routing Dispatch to Run Multiple Workers in ParallelDeepSeek 最近放出了一套 Super-agent 设计，我看完最大的感受是：以后我们可能不用再盯着 Agent 干活了。 现在用 Agent，经常会遇到一个很烦的问题。DeepSeekImageOriginal post16KLikes 150Geek11-01ZHOpenCode Go's Inference Consumes Too Much; the Official DeepSeek Interface Is Better Value- 博主实测对比发现 OpenCode Go 上模型把 51.2% 的输出花在推理上（DeepSeek API 仅 39.2%），DeepSeekOriginal post15KLikes 53
小互11-02ZHRunning Multiple Tasks in Parallel with a Router That Only Dispatches and Workers That Only Do the Work同时开 5 个 Agent，人就盯不过来了 DeepSeek Harness 团队公开了一套 Super-agent 设计，思路很直接：session 别让人管了，交给系统。 你只跟同一个 Agent 说话，它在后台同时推进好几件事。DeepSeekImageOriginal post15KLikes 89
歸藏(guizang.ai)11-03ZHInstalling a Design Skill Plugin in DeepSeek Harness and Building a Portfolio Library给我的 DeepSeek Harness 插件 dsh-guizang-design 做了个宣传片DeepSeekImageOriginal post13KLikes 108雨哥向前冲11-02ZHWhen Creator Mode Lacks a Feature, Have It Install a Plugin for Itself太炸裂了！DeepSeek 这回是把软件的遥控器直接塞到用户手里了 以前用软件，缺个功能怎么办？ 去 GitHub 提 issue，然后等。等排期，等版本，等官方心情。 GitHub 26.9 万星的 DeepSeek Harness，换了个玩法：DeepSeekImageOriginal post10KLikes 34SuSu_酥酥👅11-01ZHSteps to Run DeepSeek Harness Locally on an Android Phone🔥人气王！DeepSeek Harness 直接装进口袋，手机自己跑 Agent，再也不断档，电脑远程接着干！ 社区爆款 DSH Go 把完整引擎塞进安卓手机：内置 Ubuntu 环境免 Root 免 Termux，本机直接跑。DeepSeekImageOriginal post8.7KLikes 97
王小庄11-02ZHHands-on test of DSH super agent handling repo tickets automatically, plus three pitfalls别人都在复述 README，我把 DeepSeek 的 dsh-superagent 真装起来跑了一遍 25 分钟装完，提个 issue 54 秒它就回"在办了"，1 分 42 秒交出带回归测试的 PRDeepSeekOriginal post8.6KLikes 65SuSu_酥酥👅11-03ZHRunning a 1-Bit Quantized DeepSeek Locally on a 12GB Graphics Card😱逆天了！12GB显卡就能跑！7630亿参数压到1 bit，DeepSeek V4.1 Flash 1017 最新魔改版来了！专为代码生成 + Agent工作流优化！DeepSeekImageOriginal post8.5KLikes 122Lonely11-01ZHQuantization Plan for Running DeepSeek V4.1 Flash Locally with 64GB of RAM🎉 为何如此优秀？！Deepseek 已经进化到在 64G 内存上运行了！ @UnslothAI 再次带来黑科技，发布 DeepSeek-V4.1-Flash-1017 GGUFs !DeepSeekImageOriginal post8.1KLikes 58Kxians11-01ZHCalling DeepSeek via AMD's Free Interface, Which Can Connect to Claude Code🔥【羊毛速报】AMD 这波羊毛可以薅，Radeon Cloud 中国站免费开放了 15 款模型 API！ 包括 DeepSeek V4.1 Flash、GLM-5.3-Flash、DeepSeekOriginal post8KLikes 47雨哥向前冲11-01ZHThe Real Reasons DeepSeek Gets Slower and Dumber in OpenCode你是不是觉得 OpenCode 里的 DeepSeek 变笨了，变慢了？ 今天凌晨 OpenCode 作者出来发帖，把 7 个配置拉去跑了基准。 表格里 DeepSeek V4.1 Flash 各配置总分在 89.3 到 90.8，DeepSeekImageOriginal post6.6KLikes 25
AIGCLINK11-03ZHUsing creator mode to have DeepSeek generate the missing plugin on the spotDeepSeek越来越有意思了，你DeepSeek Harness上缺什么插件，换Creator Mode模式，直接用自然语言让AI在Harness上现给你攒 不愧DeepSeek，果然一切皆插件。DeepSeekImageOriginal post5.6KLikes 32
xiax｜AI & Agents11-02ZHCreator Mode Generates Plugins from One Sentence; Official Demo Prompt Included🚨真正的万物皆插件！彻底打破封闭AI！Deepseek Harness创作者模式开启全民造Agent时代！ ✧ 依托 Cordis 内核 Everything is a plugin 架构，UI、工具、技能全部都可插件化重组DeepSeekImageOriginal post5.5KLikes 8Penny777 🦊11-01ZHSave Money by Setting Free DeepSeek as the Main Model for Hermes好消息！WorkBuddy 的 DeepSeek 还能继续免费用。 Hermes + DeepSeek 是我最近打项目用得最多的配置了，DeepSeekOriginal post5.2KLikes 19Gorden Sun11-01ZHAutomation Agent Prototype from DeepSeek's Open-Source GitHub Repositorydsh-superagent-prototype：DeepSeek让 AI 自主管理代码仓库的实验原型 人在用 AI 写代码时，经常会被各种询问和通知打断，耗费了大量精力。DeepSeekImageOriginal post4.8KLikes 28Ren11-01ZHRunning DeepSeek Locally on a Gaming Graphics Card and Connecting It to Claude Code🎮 不敢信，FreeToken 能让一张游戏显卡跑起 DeepSeek-V4-Flash、GLM-5.2 这种大号 MoE 模型。 上过 GitHub Trending，仓库 7 月才建，现在已经 1.5 万 star，fork 也有一千五百多个。DeepSeekOriginal post4.4KLikes 81
Wei11-02ZHSpeeding Up MoE Models That Don't Fit in VRAM by 3x With an Expert Cache Patch显存装不下的 MoE，让 CPU 分担一部分专家，Qwen3.8-Flash-Next 在一张 4090 上从 29.8 tok/s 跑到了 97.3 tok/s。 这是 tandem.cpp，一组给 llama.cpp 打的补丁，刚开源。DeepSeekImageOriginal post4.2KLikes 35
oil-oil11-01ZHHow DeepSeek Builds Admin Pages: Generate an Image First, Then Reproduce ItOil UI Skill 更新了新版本，这次主要优化了用于中后台设计的能力，并且把先用 AI 生成 UI 设计稿，再用代码还原结合到了 Skill 的工作流里。 目前 Oil UI 在 GitHub 上已经有 2600 多个 Star。DeepSeekImageOriginal post4.2KLikes 46xiax｜AI & Agents11-02ZHPlan Fixed Steps Once and Skip Model Callbacks to Save Nearly 90% of Tokens🚨Token狂砍87.93%！DeepSeek Harness大瘦身！清华微软HarnessMM不换模型做强Agent ✧ 传统Agent存在明显浪费：工具反馈后反复回调LLM，编译、DeepSeekImageOriginal post4KLikes 33阿蔺A-Lin11-01ZHInstalling APKs on Android Phones Without Root Using DeepSeek Harness安卓也能跑 Agent 了，免 Root 装个 APK 就能用 DeepSeek Harness🔥 想在手机上玩 Agent，以前要么 Root，要么塞一套 Termux。DeepSeekImageOriginal post3.7KLikes 27
李岳11-03ZH7 Official Demo Cases for DeepSeek Harness Creator ModeDeepSeek Harness 刚发了条推，公布了他们的全新功能 Creator Mode（创作者模式）。 我盯着那个演示视频反复看了三遍，越看越精神。DeepSeekImageOriginal post3.7KLikes 15AI少年11-01ZHHow to Divide Work Among AI Intern Plugins in DeepSeek Harness🔥 我去，DeepSeek 把内部在用的「AI 超级同事」给开源了 开源成 DeepSeek Harness（DSH）的一个插件。 它可以给你的 GitHub 仓库招一个 AI 实习生。DeepSeekImageOriginal post3.6KLikes 25The next edition of DeepSeek Usage Tips starts on Dec 1
Covers practical Grok Bot usage from the Chinese-speaking community in Nov 2026: task assignment and multi-agent scheduling, quota consumption and ways to save quota, integration with domestic services and office tools, and lessons learned from designing Bots. Suited for people who want to use it more stably and economically.
This covers practical tips for Claude Code from the Chinese-speaking community in November 2026: intensity settings that save quota, preventing long tasks from idling, a plan-first-then-review workflow, and how to use design and chart Skills and role-division templates. Suited to those who want to use it more stably and economically.
This covers practical Codex content from the Chinese-speaking community in November 2026: ways to save tokens and quota, the quota pooling rules starting December, Skill and knowledge base configuration, and workflows for pairing with other Agents. Suited to those who want to use Codex more economically and stably.