文 | 斯凯有话说
2026年10月16日，vLLM官方博客发布了一篇技术复盘，标题很技术流：《Hyper Optimizing Kimi K3 in vLLM: Moving Fast and Responsibly》。
署名作者是英伟达CentML的工程师Benjamin Chislett。他做了一件事：用一段1.2万行CuTe DSL编写的Megakernel，将月之暗面2.8万亿参数的Kimi K3推理速度大幅提升。据SemiAnalysis测算，在相同的GB300 NVL72硬件上，K3达到423 tok/s/user的推理速度，性能提升1.68倍。vLLM博客原文报告的数据则是：batch 1下超过840 tokens/s，多个测试配置下解码吞吐提升1.5—2倍。
没有换一张卡。没有上更先进的制程。没有等下一代芯片。
但真正让这件事值得深挖的，不是那1.68倍。
是博客里的一句话，被绝大多数中文报道忽略了：
“Agent-optimized inference code is outpacing our review and maintenance workflow.”
智能体优化的推理代码，正在超过人类团队的审查和维护速度。
这句话翻译成大白话就是：AI写的代码，人类来不及审了。
一个工程师、1.2万行代码、AI辅助开发
先看技术层面发生了什么。
Kimi K3不是普通模型。它的Stable LatentMoE架构有896个专家，每次推理激活16个。这种极端稀疏的MoE，在训练端是“用更少激活参数撬动更大容量”的利器，但在推理端是工程噩梦。据SemiAnalysis分析，K3一次前向计算的HBM带宽需求约1.5TB，每次前向触发超过120次All-to-All通信。解码阶段是典型的内存带宽受限场景，计算单元大部分时间在等数据搬运。
传统推理引擎的解码过程，由一连串独立内核完成：注意力一个、MoE路由一个、专家计算一个、通信一个。每一个内核启动有开销，每一次内核间数据交换要经过显存读写。
Megakernel的核心思想是算子融合：把分散的解码内核合并成一个高度融合的超级内核，减少启动开销、合并内存访问、让计算和通信尽可能重叠。
1.2万行代码的含金量，是把“通用硬件跑通用模型”的默认效率，替换成了“专用硬件跑专用模型”的极致效率。硬件没变，模型没变，变的是“数据在硬件上怎么走”。
而写这1.2万行代码的过程，本身就是一个信号。vLLM博客明确写道：“Coding agents can now help with some of the hardest parts of model optimization: writing complex GPU kernels, performing cross-operator fusions, overlapping computation with communication, and tuning execution across different batch sizes and GPU configurations.”团队“使用了智能体辅助开发来构建统一的CuTe DSL解码路径”。
这不是“AI独立写完了全部代码”。人类工程师Benjamin Chislett是项目的主导者和架构设计者。但在最困难的部分——复杂GPU内核编写、跨算子融合、通信重叠、不同批次和GPU配置下的调优——AI承担了相当大的工作量，以至于人类的逐行审查跟不上了。
“来不及审”意味着什么
博客里还有一个细节，中文报道几乎没有提及。
vLLM团队为了应对这个局面，发布了一个fast-track model optimization RFC。核心思路不是“让人类审得更快”，而是改变性能工作的组织方式。博客原文列出了四条原则：Isolate changes（隔离变更）、Bound the scope（限定范围）、Validate correctness systematically（系统化验证正确性）、Scale validation（扩大验证）。
更重要的是这句话：“Human reviewers still judge the design and the evidence, but they can start from a high-level abstraction, instead of a line-by-line code review.”
人类审查者仍然判断设计和证据，但他们可以从高层抽象开始，而不是逐行审代码。
翻译一下：人类从“写代码的人”变成了“设计验证框架和风险边界的人”。
AI写代码，人类管流程。这不是未来，这是2026年10月16日vLLM博客里写着的现状。
这件事的分量，需要放在一个更大的背景里看。
你之前写过“AI攻击联合国网站”“AI研究员辞职警告”“AI差点引发中美冲突”。那些故事里，AI是“行为的主体”——它做了人没让它做的事。而这次不一样。这次是AI成了“优化的主体”——它在帮人写代码，让另一个AI模型跑得更快。
区别在于：前者是人类要“管住”AI，后者是人类要“跟上”AI。管住，是控制边界的问题。跟上，是能力速度的问题。
当AI生成代码的速度，超过了人类审查代码的速度，“人类在回路中”这个原则，就面临一个非常具体的挑战：回路上的人类，还能不能看懂回路里的东西？
为什么解码优化比训练优化更难
要理解这1.2万行代码为什么值钱，得先理解K3的推理为什么难。
训练是“算力峰值受限”——你有多少FLOPs，就能跑多少计算。解码是“内存带宽受限”——计算单元大部分时间在等数据搬运。K3的896个专家被分布到多张芯片上，每处理一个token，路由系统要把数据分发到被选中的专家所在芯片，再把结果收回。超过120次All-to-All通信，每一次都是延迟。
所以在这个场景里，“换更快的芯片”不一定是最高效的解法。瓶颈不在算力峰值，在数据流动的效率。Megakernel的算子融合，本质上是在减少数据在显存和计算单元之间来回搬运的次数。
这个技术判断的产业含义是：当制程红利放缓，当单卡峰值性能追赶需要以年为单位计算，“让已有硬件跑出更高有效算力”正在成为回报率更高的路径。
编译器栈，正在成为算力竞争的新前线
SemiAnalysis创始人迪伦在红杉专访中有一个判断：“性能提升绝不能单靠硬件堆砌。模型架构、软件栈（编译器与通信库）与底层硬件的深度协同优化，才是实现百倍效率飞跃的真正突破口。”
这句话对理解英伟达这1.2万行代码的意义，比任何单一的性能数字都重要。
在先进制程芯片获取受限的背景下，单卡性能差距短期难以抹平。但系统效率、编译器优化、算子融合、通信重叠——这些是纯软件层面的工程能力，不依赖晶圆厂，依赖工程师和工具链。英伟达用一个工程师、1.2万行代码、AI辅助开发，在不换硬件的前提下拿到了显著的效率提升。这个事实本身就在提示：算力竞赛的维度正在分裂。一条线是“造出更强的芯片”，另一条线是“让现有的芯片干更多的活”。后一条线，对追赶者门槛更低，空间更大。
IT之家的报道提到，英伟达“还在为vLLM开发DeepSeek V4.1等其他模型的优化”。这意味着Kimi K3的优化不是一次性的炫技，是一套可复用的方法论正在向更多中国模型扩散。
英伟达在帮中国模型“榨干”它自己卖的硬件
这里有一层微妙的关系值得点一句。
英伟达优化的是月之暗面的Kimi K3，还在优化DeepSeek V4.1。在芯片管制的背景下，英伟达的软件团队正在为中国模型“榨干”它自己卖的硬件。
卖铲子的人，在帮你磨铲子。这既是因为软件优化能拉动硬件销售，也是因为英伟达需要证明：即使在管制框架下，它依然是中国AI产业不可或缺的一环。这个关系里，竞争与合作是同时存在的。
写在最后
回到那句“来不及审”。
这件事真正的分量，不在于英伟达优化了Kimi，不在于1.68倍，甚至不在于“编译器栈成为新前线”。
在于它提出了一个人类必须回答、但目前还没有答案的问题：当AI开始优化AI，而且优化的速度超过了人类理解的速度时，人类还能不能守住“最终判断权”这条底线？
vLLM团队的答案是“改变组织方式”——不逐行审，而是设计验证框架、限定变更范围、系统化测试。博客里还披露了一个细节：审计发现编译器把激活读取移到了同步等待之前，修复方案是把这种加载类型限制为静态权重。这类错误能通过简单的功能测试，需要针对性的调查才能发现。这个答案务实，但它本质上是在说：我们接受“看不懂每一行代码”这个现实，然后用流程来兜底。
这可能就是AI时代“人在回路中”的新形态。不是“人看懂一切”，而是“人定义边界、设计验证、承担后果”。
技术可以加速，但判断不能外包。这句话在“AI写代码优化AI”的场景里，有了一个新的含义：判断不一定看得懂代码，但必须看得住风险。
vLLM的博客标题里有“Responsibly”这个词。在一个AI生成代码速度超过人类审查速度的时代，“负责任”这三个字的分量，比1.68倍重得多。
AI开始自主优化AI、代码迭代速度碾压人类审核能力，算力竞争正式从「拼硬件」迈入「拼软件栈、拼人机治理」的新时代。未来不会是人类淘汰AI，而是跟不上AI迭代节奏的人，会被行业淘汰。
你认为：当AI写代码的速度超过人类审查的速度，你觉得“人在回路中”还守得住吗？人类放弃逐行审代码、靠框架兜底风险，真的能守住AI的最终控制权吗？关注我，评论区聊聊你的看法。
免责声明
本文内容基于vLLM官方博客、SemiAnalysis及公开行业报道整理，文中性能数据、技术细节以原始信源为准。文中所有分析、推论与观点仅为个人产业观察，不构成投资建议、商业决策或政策解读依据。
行业技术处于快速迭代状态，信息存在滞后、偏差可能性；读者依据本文内容作出任何商业、投资相关判断，需自行承担全部风险。
本文无商业赞助，仅作行业交流。转载请注明出处。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Yan, Skye has something to say.
On October 16, 2026, vLM official blog publish wrote a technical replay entitled: Hyper Optimizing Kimi K3 in vLM: Moving Fast and Responsibly.
The author is Benjamin Chislett, an engineer at CentML, Inc., Inc. He did one thing: Megakennel, written in a 12,000-year line of Cute DSL, significantly raised Kimi K3 ' s reasoning on the dark side of the moon by 2.8 trillion parameters. On the same GB300 NVL72 hardware, K3 was calculated to be 423 tok/s/user ' s reasoning speed, with a performance increase of 1.68 times. The data reported in the original version of the vLM blog are over 840 tokens/s for the bat, and 1.5 - 2 times higher for multiple test configurations.
No new card. No more advanced programming. No waiting for the next generation of chips.
But it's not 1.68 times what really makes it worth digging.
It's a word in a blog that has been ignored by most Chinese reports:
“Agent-optimized inference code is outpacing our review and maintenance workflow.”
The optimal code of reasoning for AI agent is exceeding the speed of review and maintenance by the human team.
And that's translated into the big white saying: "Ai wrote the code, man can't read it."
One engineer, 12,000 line code, AI-aided development
Let's see what happens at the technical level.
Kimi K3 is not an ordinary model. Its Stable LatenmoE structure has 896 experts, 16 of whom are activated each time the reasoning is made. This extremely rare MoE, at the training end, is a sharp instrument “to pry a larger capacity with fewer activated parameters”, but at the reasoning end is an engineering nightmare. According to SemiAnalysis analysis, K3 had a forward-calculated HBM bandwidth requirement of approximately 1.5 TB, triggering more than 120 All-to-All communications each. The decodering phase is a typical memory bandwidth-limited scenario, with the computing unit waiting for data removal for most of the time.
The decoding process of the traditional reasoning engine was carried out by a series of independent kernels: one focus, one MoE road, one expert, one communication. Each kernel launch has costs and each kernel data exchange is to be read and written through VRAM.
Megakernel's core idea is the integration of algorithms: the consolidation of decentralized decoded inner cores into a highly integrated superkernel, reduced start-up costs, consolidated memory access, and as much overlap as possible between computing and communications.
The gold content of line 12,000 codes is the default efficiency of the General Hardware Run General Model, which is replaced by the extreme efficiency of the Specialized Hardware Run Model. The hardware remained the same, the model remained the same, and the change was “how data go on hardware”.
And the process of writing this 12,000-line code is itself a signal. The vLM blog clearly states: “Coding events can now help with some of the best parts of Model Optification: working complex GPU processes, working cross-purposes, overlaping communications with communization, and going forward accounting transactions differing nets and GPU decodings.”
This is not "AI wrote all the codes independently." Benjamin Chislett, an human engineer, is the lead and architecture designer of the project. But in the most difficult part — complex intra-GPU writing, cross-calculator integration, overlapping communications, different batches and fine-tuning under the GPU configuration — AI has taken on a considerable amount of work, so that the human line-by-line review has not kept pace.
What does "failure" mean?
There's one more detail in the blog, and there's hardly any mention in Chinese.
In response to this situation, the vLM team publish has a fast-track model optimation RFC. The core idea is not “to make human beings judge faster”, but to change the way performance works. Four principles are listed in the original blog: Isolate Changes (Isolation Change), Bound the scope (limited), Validate Corporation System (Systemized Validity Validation) and Scale Validation (Enlarged Validation).
"Human reports still judge the declaration and the appeal, but they can start from a high-level intervention, instead of a line-by-line code review."
Human reviewers continue to judge design and evidence, but they can begin with high-level abstraction rather than line-by-line codes.
Translating: Human beings have moved from “the person who writes the code” to “the person who designs the verification framework and the risk boundary”.
AI writes codes, humans handle processes. It's not the future, it's the status quo on the VLM blog on October 16, 2026.
The weight of the matter needs to be seen in a larger context.
You wrote "AI attacks the United Nations website" "AI Fellow's Warning to Resign" "AI almost triggered a Central American conflict." In those stories, AI is the "subject of behavior" -- it does what people don't let it do. And this is different. This time, AI became an "optimal subject" -- it's helping people write codes, making another AI model run faster.
The difference is that the former is the human being who “takes over” AI, while the latter is the human person who “takes up” AI. Take care, it's border control. Keep up, it's power speed.
When AI generates codes faster than humans review codes, the principle of “humans in return” faces a very specific challenge: are humans on their way back, and can they understand what goes back?
Why is it harder to decode it than to train it?
It is difficult to understand why the 12,000-line code is worth money.
The training is "the limit of the calculus peak" -- you can run as many FLOPS as you can. Decoding is "RAM bandwidth limitation" - the computing unit is waiting for data removal for most of the time. The 896 K3 experts were distributed over a number of chips, each token processing, and the route system was to distribute the data to the chip of the selected expert for recovery. More than 120 All-to-All communications were delayed each time.
So in this scenario, changing the faster chip is not necessarily the most efficient solution. The bottlenecks are not at the peak of power, but at the efficiency of data flows. Megakernel's algorithms are integrated in essence to reduce the number of round-trip movements of data between VRAM and the computing unit.
The industry in this technical sense is that when the coupling dividend slows, when single-cardo peak performance takes place in units of years, “making available hardware more effective” is becoming the path to higher returns.
The compiler's vault is becoming a new front for computing.
The founder of SemiAnalysis, Dylan, made a judgement in his interview with Redwood: “Performance must not be built on hardware. Model architecture, software stacks (compilers and communication libraries) and depth optimization of bottom hardware are the real breakthroughs in achieving 100-fold efficiency leaps.”
This sentence is more important than any single performance figure to understand the meaning of the 12,000-line code in Weida.
In the context of limited access to advanced chip production, single card performance gaps are difficult to close in the short term. But system efficiency, compiler optimization, algorithm integration, communication overlap - These are purely software-level engineering capabilities, not dependent on the crystal mill, but on engineers and tool chains. With an engineer, 12,000 line code, AI-aided development, YVD achieved significant efficiency gains without changing hardware. The fact itself suggests that the dimensions of the calculus are being divided. One line is “making stronger chips” and the other line is “making existing chips do more”. In the latter line, the threshold for chasers is lower and the space is greater.
According to a report from IT House, Inweida “is also developing the optimization of other models such as DeepSeek V4.1 for vLM”. This means that Kimi K3 ' s optimization is not a one-time showdown, but a reusable methodology that is spreading to more Chinese models.
Young Wai Da is helping the Chinese model dry up its own hardware.
There is a delicate relationship worth saying.
Young Weidar optimised Kimi K3 on the dark side of the moon and is also optimizing DeepSeek V4.1. In the context of chip control, the software team in Inweida is “driven” its own hardware for the Chinese model.
Selling shovels is helping you grind shovels. This is due both to the fact that software optimization pulls hardware sales and to the fact that Yin Weida needs to prove that even under a regulatory framework, it remains an essential part of China ' s AI industry. In this relationship, competition and cooperation coexist.
At the end.
Go back to the phrase "not in time for trial."
The real weight of this matter is not because Ingwida optimized Kimi, or 1.68 times more, or even “the compiler vault became the new front line”.
It is that it raises a question that human beings must answer, but which they have not yet answered: can human beings hold the bottom line of the “final judgement” when AI begins to optimize AI and at a rate that exceeds human understanding?
The answer of the vLM team is “change organizational style” - not on a case-by-case basis, but on the design of the validation framework, limit the scope of the change, and systematized testing. A detail was also disclosed in the blog: the audit found that the compiler had moved the activation reading to the synchronized waiting period, and that the restoration programme limited this load type to static weights. Such errors can be detected through simple functional tests that require targeted investigations. This is a pragmatic answer, but it's essentially saying: We accept the reality that we cannot read every line of codes, and then we use the process to go to the bottom.
This may be the new form of the "people in return" of the AI era. Not “people understand everything”, but “people define boundaries, design certification, bear consequences”.
Technology can accelerate, but judgement cannot be outsourced. This phrase has a new meaning in the context of "AI Write Code Optimize AI": judgements do not necessarily understand codes, but they must be risk-sensitive.
The title of the vLM blog contains the word “Responsibly”. In an age when AI generation codes are faster than human scrutiny, the word “accountability” weighs more than 1.68 times.
AI began to independently optimize AI, code-overtone speed, crushing human auditing capabilities, and computing competition formally moved from "hardware puzzles" to a new era of "scrolling of software stacks, plaster governance." The future will not be for humans to phase out AI, but for those who fail to keep up with the acoustic pace, the industry will be eliminated.
Do you think that when AI writes codes faster than human scrutiny, you think that "people on the way back" can hold? Is it true that the ultimate control of AI can be secured by humans abandoning the line-by-line code and using frame-by-frame risk? Watch me, comment section.
Disclaimer
The content of this paper is based on v LLM official blogs, SemiAnalysis and open industry reports, with performance data and technical details based on original sources. All the analyses, inferences and views contained in the paper are personal industry observations and do not constitute a basis for investment advice, business decision-making or policy interpretation.
Industry technology is in a state of rapid iterativeity and information is subject to lags and deviations. Readers are required to take full responsibility for any business and investment-related judgement based on the content of this paper.
This paper is free of commercial sponsorship and only for industry exchanges. Please indicate where you are.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.