白宫指控中国AI公司"偷技术"，中国商务部回击"AI霸权主义"。Kimi K3发布一周，全球都在吵"蒸馏"。

大部分吵这个的人，根本不知道蒸馏是什么。

这篇文章用最直白的话说清楚：蒸馏是技术还是盗窃？为什么专家说"纯蒸馏做不出Kimi K3"？以及——为什么真正让美国急的不是"偷"，而是开源模型让闭源API卖不动了。

蒸馏到底是个什么东西

先说清楚：蒸馏不是一个藏在暗处的黑客手段。它是AI行业公开使用的一种训练技术，有论文、有标准实现、在顶级会议上发表过。

工作原理很简单：

训练大模型很贵。GPT-5的训练成本估计数亿美元。但用一个已经训练好的模型很便宜——调一次API几美分。

蒸馏的思路是：让一个小模型反复调用一个大模型的API，把大模型的输出当作"标准答案"来学习。小模型学会了大模型的"思维方式"，然后你关掉大模型，小模型自己就能干差不多的活。

一个类比：

这不是复印学霸的作业本。这是坐在学霸旁边，看他做了一千道题，然后自己学会了怎么解题。

你抄作业本，是1:1复制，那是侵权。你学会了解题思路，那是学习。

蒸馏在技术层面的真相更接近后者——但又不完全是。因为模型"学习"的方式比较机械：它没法像人一样"理解"，它只能大量统计规律。但在结果上，它确实是在"学会"能力而不是"复制"数据。

行业惯例：

蒸馏不是中国公司的发明。硅谷每天都在用蒸馏：

  • 谷歌用蒸馏训练Gemini Nano（手机版小模型）
  • OpenAI用GPT-4蒸馏过GPT-4 Turbo
  • Anthropic自己的Fable模型训练过程中也用了蒸馏技术
  • 开源社区的Alpaca、Vicuna都是蒸馏出来的成果

2023年斯坦福的Alpaca项目——用Meta的LLaMA蒸馏OpenAI的GPT-3.5——当时被当成Open Source的典范来报道，没人说这是"偷"。

为什么专家说Kimi K3不是蒸馏出来的

这是争议的核心。白宫科技政策办公室主任Kratsios说月之暗面"蒸馏了Anthropic的Fable"。但多个独立专家（包括OpenAI自己的总裁Brockman）都承认：K3不太可能是蒸馏做出来的。

原因有三：

1. 时间窗口不够。

Kimi K3于7月17日发布。Anthropic的Fable 5是6月下旬才对外开放API的。从Fable开放到K3发布，中间不到三周。

蒸馏不是一两天的事。要做工业级别的蒸馏，你需要：

  • 数百万次API调用（采集训练数据）
  • 数据处理和清洗
  • 小模型的训练、调参、验证
  • 多轮迭代

三周时间做这些，技术上极不现实。更关键的是——K3有一些能力是Fable不具备的（比如在Frontend Code Arena上超过Fable排第一），如果它只是蒸馏Fable，不可能在Fable擅长的领域超过Fable。

2. K3有底层原创架构。

月之暗面发布了技术报告，公开了三项自研架构：

  • Moon Clip：二阶优化器，让20T数据的训练效果接近40T
  • Kimi Linear Tension：线性注意力机制，支持100万token上下文
  • Attention Residuals：注意力残差，推理效率提升25%

这些不是蒸馏能拿到的。蒸馏只能学习模型的外部行为，学不到内部架构。如果K3有这些独特的架构设计，说明它从底层就是自己造的。

3. 闭源模型很难被有效蒸馏。

Anthropic的Fable 5是闭源模型。你只能通过API调用它，看不到它的权重、架构细节、训练数据。蒸馏一个闭源模型，你能得到的最好结果是"模仿者"，不可能达到甚至超过原模型。

这就好比你每天看梅西踢球，你学会了他的一些跑位和技巧，但你不可能因此变成梅西。

合法蒸馏 vs IP盗窃：边界在哪里？

那美国在急什么？

合法的蒸馏： 开源模型之间互相学习、小团队蒸馏公开API来加速研究、学术界的知识蒸馏研究——这些全球通用，没问题。

被认为越界的蒸馏（"蒸馏攻击"）：

  • 以工业规模调用API（每天数百万次），专门为了提取教师模型的能力
  • 违反API服务条款（大多数AI公司的ToS明确禁止用于训练竞争模型）
  • 用量大到你实际上是在建立一个"平行副本"

白宫指控月之暗面做的是第二种。问题是——目前没有公开证据。

Anthropic没有公布任何数据证明K3蒸馏了Fable。Kratsios的指控是基于"我们掌握信息"的笼统说法，没有技术细节。

更有意思的矛盾：Anthropic上个月刚指控阿里巴巴蒸馏了它的Claude模型，现在又指控月之暗面蒸馏了Fable。如果月之暗面也蒸馏了、阿里也蒸馏了、DeepSeek也蒸馏了——那中国这家叫蒸馏、美国公司自己做的那个叫"优化"？

双标是人类的本性，但它经不起推敲。

真正让美国急的，不是"偷"

把"蒸馏"争议拆干净之后，你会发现核心矛盾其实很简单：

不是因为中国AI偷了美国技术——而是因为中国的开源模型正在改变行业的经济模型。

Kimi K3的API定价大约是Anthropic Fable的15-20%。

微软正在评估用K3替换部分原本由GPT和Claude处理的任务——因为便宜、够用、能自己部署。

如果越来越多的企业和开发者选择中国开源模型，美国的闭源API商业模式就面临根本性挑战：你融资650亿美元做出来的模型，它的市场价值被一个20亿美元融资的中国开源项目侵蚀。

所以问题不是"中国AI偷了美国什么"，而是"中国AI正在证明顶尖模型不一定需要天价融资和闭源收费"。

芯片可以禁运，先进GPU可以在海关拦住。但模型权重——2.8万亿个参数，压缩成几百GB的文件，挂在GitHub上——你拦不住。

黄仁勳最近的原话： "将英伟达芯片挡在中国市场之外，已经不再意味着能够把中国挡在AI竞争之外。"

他说的就是这个意思。

写在最后

蒸馏争议还会继续。美国政府显然在推动对中国开源的限制，但这条路比芯片禁运难走得多——因为开源模型不存在"海关"。

但技术层面的事是清楚的：

  1. 蒸馏是一个通用的AI训练技术，不是盗窃
  2. Kimi K3的技术水平和发布节奏，不支持"纯蒸馏"的指控
  3. 真正驱动这场争议的，不是技术窃取的担忧，而是开源对闭源商业模式的威胁

下次再看到有人说"中国AI偷了美国的技术"——你可以问问他们，什么叫蒸馏。

The Truth About AI Distillation: It's Not What You Think

The White House accuses Chinese AI companies of "stealing technology." China's Ministry of Commerce fires back with accusations of "AI hegemonism." Since Kimi K3 dropped, the whole world has been arguing about "distillation."

Most people arguing about it have no idea what distillation actually is.

This article explains, in plain language: Is distillation a legitimate technique or theft? Why do experts say "you can't build Kimi K3 with pure distillation"? And — the real reason the US is worried isn't "theft," it's that open-source models are disrupting the closed-API business model.

What Distillation Actually Is

Let's be clear from the start: distillation isn't some shady hacker trick. It's a standard training technique used openly across the AI industry, with published papers, reference implementations, and a track record at top conferences.

How it works:

Training a large model is expensive. GPT-5 reportedly cost hundreds of millions in training compute. But using a trained model is cheap — a few cents per API call.

The idea behind distillation: have a small model repeatedly query a large model's API, using the large model's outputs as "ground truth" training data. The small model learns the large model's "way of thinking." Then, you can shut down the expensive large model and let the small one handle the same tasks at a fraction of the cost.

An analogy:

This isn't photocopying the smart kid's homework. This is sitting next to the smart kid, watching them solve a thousand problems, and then learning to solve problems yourself.

Photocopying is 1:1 reproduction — that's infringement. Learning how to solve problems — that's learning.

Distillation operates closer to the latter, technically. But not exactly — because models don't "understand" like humans do. They learn statistical patterns, not concepts. In outcomes though, they genuinely acquire capabilities, not copy data.

Industry standard:

Distillation isn't a Chinese invention. Silicon Valley uses it daily:

  • Google distills Gemini into Gemini Nano (the phone-sized model)
  • OpenAI distilled GPT-4 into GPT-4 Turbo
  • Anthropic used distillation during Fable's own training
  • The open-source Alpaca and Vicuna projects (2023) were distillation products

When Stanford released Alpaca in 2023 — distilling Meta's LLaMA from OpenAI's GPT-3.5 — it was celebrated as an open-source milestone. Nobody called it "theft."

Why Experts Say Kimi K3 Isn't Distilled

This is the core of the controversy. White House tech advisor Kratsios claims Moonshot "distilled Anthropic's Fable." But multiple independent experts — including OpenAI's own president Greg Brockman — acknowledge that K3 was unlikely built through distillation alone.

Three reasons:

1. The timeline doesn't work.

Kimi K3 launched July 17. Anthropic's Fable 5 opened its API in late June. That's less than three weeks.

Real distillation at scale requires:

  • Millions of API calls to collect training data
  • Data processing and cleaning
  • Student model training, tuning, and validation
  • Multiple iteration rounds

Three weeks is technically infeasible. More importantly — K3 outperforms Fable on certain benchmarks (Frontend Code Arena #1). If it were a pure distillation of Fable, it couldn't exceed Fable on Fable's own strengths.

2. K3 has original architecture.

Moonshot published a technical report detailing three self-developed innovations:

  • Moon Clip: a second-order optimizer achieving 20T-data effectiveness comparable to 40T
  • Kimi Linear Tension: linear attention mechanism supporting 1M token context
  • Attention Residuals: improving reasoning efficiency by 25%

Distillation only captures external behavior, not internal architecture. If K3 has unique design innovations, that proves it's built from the ground up, not copied.

3. You can't effectively distill a closed model.

Fable 5 is closed-source. You can only access it through an API — no weights, no architecture details, no training data. Distilling a closed model gives you an imitation at best. You can't match or exceed the original.

It's like watching Messi play every day. You'll learn some movement patterns. You won't become Messi.

So what's the US actually upset about?

Legal distillation: open-source models learning from each other, small teams distilling public APIs for research, academic knowledge distillation research — these are global standard practice, no problem.

What's considered over the line ("distillation attack"):

  • Industrial-scale API calling (millions per day) specifically to extract a teacher model's capabilities
  • Violating API terms of service (most AI companies' ToS explicitly prohibit using their API to train competing models)
  • Building a de facto parallel copy through sheer query volume

The White House accuses Moonshot of the second category. Problem is — there's no public evidence.

Anthropic hasn't released any data proving K3 distilled Fable. Kratsios's accusation is based on vague "we have information" language with zero technical detail.

The irony: Anthropic just accused Alibaba of distilling Claude last month, and now they're accusing Moonshot of distilling Fable. If Alibaba did it, Moonshot did it, DeepSeek did it — then when Chinese companies do it, it's "theft," but when American companies do it, it's "optimization"?

Double standards are human nature. But they don't hold up to scrutiny.

What's Really Bothering the US

Strip away the distillation rhetoric, and the core issue becomes clear:

It's not about China stealing US technology. It's about Chinese open-source models changing the industry's economics.

Kimi K3's API pricing is roughly 15-20% of Anthropic Fable's.

Microsoft is evaluating K3 to replace some tasks currently handled by GPT and Claude — because it's cheaper, good enough, and deployable on your own infrastructure.

If more enterprises and developers choose Chinese open-source models, the US closed-API business model faces an existential challenge: you raised $65 billion, but your model's market value is being eroded by a Chinese open-source project that raised $2 billion.

So the real question isn't "what did China steal from the US?" It's "what if world-class AI doesn't require billion-dollar funding and closed-source pricing?"

Chips can be embargoed. Advanced GPUs can be stopped at customs. But model weights — 2.8 trillion parameters, compressed into a few hundred GB file, sitting on GitHub — you can't stop them.

Huang Renxun's recent quote: "Blocking NVIDIA chips from the Chinese market no longer means blocking China from the AI race."

That's the point.

Bottom Line

The distillation debate will continue. The US government is clearly pushing to restrict Chinese open-source AI — but this path is far harder than chip embargoes, because open-source models don't have a "customs checkpoint."

But the technical facts are clear:

  1. Distillation is a general AI training technique, not theft
  2. Kimi K3's capability and timeline don't support a "pure distillation" accusation
  3. What's really driving this controversy isn't concern about IP theft — it's the threat open-source poses to the closed-source business model

Next time someone says "China stole US AI technology" — ask them if they know what distillation actually is.