ChinAI #369：老板要我部署 Kimi K3，我该怎么办？

来自这样一个世界的问候……

蝉杀手黄蜂又回来了

**背景：**这篇文章的内容密度低于我们通常的选题，但它很好地解释了：为什么你无法在自己的设备上运行 Moonshot AI 最新的 K3 模型（中文原文）。一位朋友问快刀青衣（一款热门中文知识分享应用的联合创始人）：“我听说 K3 是一个开源且免费的模型，对吧？那么，我是不是可以直接下载它，部署在自己的电脑上，以后就再也不用花钱了？”

核心要点：与开源软件不同，“开源”语言模型通常只开放权重，而不开放源代码。

• 仅从加载 K3 所需的硬件成本来看，至少需要 16 张 H200 GPU。这与 Kimi K2 不同：爱好者曾成功在 Mac Studio 上运行其压缩版本。但如今，K3 是一种“只能运行在数据中心或企业环境中的物种”。

• 文章解释道：

“让我打个比方。想象一家米其林三星餐厅免费公开了其招牌菜的完整食谱：食材用量、烹饪时间和每一个步骤都写得清清楚楚。你能复现它吗？看起来似乎可以。

直到你翻到第一页，看到厨房要求：你需要 64 台同时运行的专业级炉灶、一套足以支持小型工厂的供电系统，以及仅购买设备就需要 2000 万元人民币的初始投入。食谱确实免费，但厨房也确实贵得令人负担不起。”

我们甚至还没有计算电费。

• Moonshot AI 的官方建议是，在至少配备 64 张加速卡的超节点上部署 K3（总成本约为 1700 万元人民币）。但如果你尝试在家中让这些卡满负载运行，其功耗将达到 45 千瓦，远远超出普通家庭电表和线路所能承受的范围。

• 一些观察者指出，K3“非常消耗 token”。文章指出：“传统软件的经济学逻辑是：开发成本高昂，而复制和运行几乎免费。相比之下，大模型的经济学逻辑是：开发成本极高；虽然复制免费，但每一次运行都会以电力和 GPU 使用的形式消耗真金白银。”

全文翻译：老板让我部署免费的 K3 模型：我先算了一下“负担不起”的服务器成本

以下是我从 K3 技术报告中注意到的一些要点：

• K3 落后于两套最强的专有系统（Claude Fable 5 和 GPT-5.6 Sol），但位于成本效率前沿。

• 在后训练和评估阶段，Moonshot AI 使用了一套名为 AgentENV 的新型沙盒基础设施；该基础设施本身也是一个开源平台。

• 这表明 Kimi 团队与清华大学 MADSys Lab（机器学习、AI 与大数据系统研究组）之间存在紧密合作。该实验室还开发了 Mooncake 架构，帮助 K3 实现更高的缓存命中率。我查看 AgentEnv 的贡献者列表时发现，一名清华大学博士生的提交次数位居第一。

• K3 的网络安全能力出人意料地薄弱。它确实能够识别此前未知的漏洞，但并不擅长将漏洞转化为可用的攻击利用程序。报告中还提到：

• K3 报告披露，K3 接受了英国 AI 安全研究所和 NIST AI 标准与创新中心的独立联合评估。根据这份评估：“**在英国 AISI / CAISI 进行的初步网络安全评估中，Kimi K3 的表现显著低于最新的前沿网络安全能力模型。**具体而言：在被要求开发攻击利用程序时，Kimi K3 的表现显著低于最新的前沿网络安全能力模型。”

Lavender Au 为 The Dial 撰写的这篇长文，探讨了中国的短剧（每集仅数分钟）。仅在今年第一季度，就有 12.2 万部 AI 生成短剧上线。Au 写道：

在中国的任何一天，大约有 2.15 亿人花费超过一小时观看短剧。这些短剧在中国互联网中迅速繁殖，2025 年发布数量达到 3.3 万部。国内市场规模已超过 1000 亿元人民币（138 亿美元），是 2024 年的两倍。

Zvi Mowshowitz 汇总了 K3 在众多公开和私有基准测试中的一系列精彩判断。关于 Moonshot 是否有意“削弱”K3 网络安全能力的讨论尤其有趣。

美国国土安全部的这项新规，将使国际学生赴美求学更加困难。NAFSA：国际教育工作者协会 CEO Fanta Aw 表示：“尽管包括 NAFSA 在内的各方提交了超过 2 万条公众意见，并对该规则提出严重关切，但 DHS 仍决定推进这项规定。它会为全球人才制造更多障碍，却不会让我们的国家更安全或更强大。”

这些是 Jeff Ding 对中文 AI 及相关议题思考的（不定期）每周译文。Jeff 是乔治·华盛顿大学政治学助理教授。

Greetings from a world where…

cicada killer wasps are back

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian /Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Context: Less dense than our usual fare, this article (link to original Chinese) is a nice explainer of why you can’t run Moonshot AI’s latest model K3 on your own device. A friend asked 快刀青衣 (co-founder of a popular Chinese knowledge-sharing app): “ I heard K3 is an open-source and free model, right? So, can I just download it, deploy it on my own computer, and never have to spend money on it again? ”

Key Takeaways: “Open-source” language models differ from open-source software in that the former only open-sources the weights, not the source code.

• Based on hardware costs to just load K3, you need at least 16 H200 GPUs. This is different from Kimi K2, a compressed version of which enthusiasts managed to run on a Mac Studio. Now, however, K3 is a “species that can only run in data centers or within enterprise environments.”

• The article explains:

“Let me use an analogy. Imagine a three-star Michelin restaurant releasing the full recipe for its signature dish for free: ingredient quantities, cooking times, and every step is clearly written out. Could you recreate it? It looks like you could.

That is, until you turn to the first page and see the kitchen requirements: you need 64 professional-grade stoves running simultaneously, a power supply capable of supporting a small factory, and an initial investment of 20 million (RMB) just to acquire the equipment. The recipe is truly free, but the kitchen is truly unaffordable.”

We haven’t even mentioned the electricity costs.

• Moonshot AI’s official recommendation is to deploy K3 on a super-node of at least 64 accelerator cards (roughly 17 million RMB in total costs). But if you tried to run these cards at full load in your home, it would consume 45 kilowatts, which would far exceed an ordinary household’s electrical meter and wiring.

• Some observers have noted that K3 is “hungry for tokens.” The article points out, “ The economics of traditional software are such that development is expensive, while replication and operation are virtually free. In contrast, the economics of large models involve extremely high development costs; while replication is free, every single operation burns through real money in the form of electricity and GPU usage.”

FULL TRANSLATION:My Boss Asked Me to Deploy the Free K3 Model: I first calculated the “unaffordable” server costs

Here are some notes on things that stood out to me from K3’s technical report:

• K3 trails the two strongest proprietary systems (Claude Fable 5 and GPT-5.6 Sol) but sits at the cost-efficiency frontier.

• For the post-training and evaluation phase, Moonshot AI used a new sandbox infrastructure called AgentENV , which itself is an open-source platform.

• This speaks to the close collaboration between the Kimi team and Tsinghua University’s MADSys Lab (machine learning, AI, big data systems group). This lab also developed the Mooncake architecture that helped K3 achieve higher cache hit rates. When I looked at the AgentEnv contributors list, a Tsinghua PhD student ranks as top in commits.

• K3’s cybersecurity capabilities were surprisingly weak. It does identify previously unknown vulnerabilities but it was not very strong at converting vulnerabilities into working exploits. Also from the report:

• The K3 report disclosed that K3 underwent an independent joint assessment by the UK AI Security Institute and NIST’s Center fro AI Standards and Innovation. From the assessment : “Kimi K3 performs significantly below the most recent frontier cyber-capable models on preliminary cyber evaluations run by UK AISI / CAISI. Specifically: When tasked to develop exploits, Kimi K3 performs significantly below the most recent frontier cyber-capable models.”

By Lavender Au, for The Dial , this longform article explores China’s short dramas (which run just a few minutes per episode). In just the first quarter of this year, 122,000 AI-generated short dramas went online. Au writes:

On any given day in China, roughly 215 million people spend over an hour watching short dramas. These dramas have multiplied on China’s internet, with 33,000 released in 2025. It’s a more than 100-billion-yuan ($13.8 billion) market domestically, double what it was in 2024.

Zvi Mowshowitz rounded up a great array of judgements on K3’s capabilities across many public and private benchmarks. The discussion about whether Moonshot intentionally “nerfed” K3’s cyber capabilities was especially interesting.

This new Department of Homeland Security rule will make it even harder for international students to study in the states. From NAFSA: Association of International Educators CEO Fanta Aw: “ Despite more than 20,000 public comments—including NAFSA’s—that raised serious concerns, DHS has chosen to move forward with a rule that will create more barriers for global talent without making our nation safer or stronger.”

These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian /Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99