Inside Kimi: 100 Hours of Observation Liu Mo 2026-07-22 79.1 KB 645.3K 1.1K ☆ 收藏 𝕏 文章列表 本文讲述了作者深入 AI 独角兽 Moonshot AI（Kimi 母公司）内部进行 100 小时的观察记录。文章回顾了公司从早期的营销挣扎到面对竞争对手 DeepSeek 突围时的战略调整，展现了这家年轻、高估值且充满内向天才员工的神秘企业文化。 KimiDeepSeekLLM人物企业文化创业人工智能月之暗面观察报道 # 100 Hours Inside Kimi **作者**: Rui Ma **日期**: 2026-04-01T07:38:24.000Z **来源**: [https://x.com/ruima/status/2039245985520681257](https://x.com/ruima/status/2039245985520681257) ---  This is a translated article from Chinese, originally published by Renwu, a respected Chinese magazine known for reported features and profile writing. It has been adapted for readers who may not know the Chinese context, company culture, or references. I’ve made it more readable, added brief context where needed, and smoothed some phrases that would sound strange if translated directly. By Liu Mo Edited by Jin Zha Originally published in Chinese by Renwu (人物) on March 31, 2026 Spring 2026 has been unusually kind to Kimi. In just a few months, the company behind Kimi seemed to hit one milestone after another. Its revenue, fundraising, and valuation all kept breaking records. A research paper co-authored by a 17-year-old high school intern received praise from Silicon Valley figures including Elon Musk. And Cursor, the U.S. coding startup valued at around $50 billion, was accused by Chinese observers of essentially “wrapping” or heavily relying on Kimi’s model as part of its own product experience. In other words, Kimi suddenly seemed to be winning on all three fronts at once: capital, technology, and commercial traction. This startup is only three years old. Its valuation has already surpassed RMB 120 billion, or roughly $16 billion. It is becoming impossible to ignore in the global AI story. And yet Moonshot AI, the company behind Kimi, remains deeply mysterious. I was given permission to spend 100 hours observing the company from the inside. As an independent writer, I was allowed to interview any employee willing to talk, sit in on any meeting that did not involve trade secrets, and write freely afterward. No one would edit my work. I would not be paid. That, it turns out, is very much in character for this company. Inside the office, it feels like standing in the eye of a storm. At the center, everything is strangely still. The desks are quiet. Only scattered keyboard sounds break the silence. Occasionally you hear someone laugh. But the noise outside, the rumors, arguments, hype, imitation, and endless commentary, seems to leave no trace here. There are just over 300 employees. Their average age is under 30. Each person, if you divide the company valuation by headcount, is effectively carrying close to RMB 400 million in enterprise value on their shoulders. About 80% of the staff are what Chinese internet slang calls “I people,” meaning introverts, borrowing from MBTI language. People sit side by side, but they are more comfortable typing than talking. Here, introversion is not treated as a flaw. It is almost an operating protocol. I thought back to my first visit in 2024, on a night when the storm was only beginning to gather. At the time, I did not come away with a particularly positive first impression. ## “DeepSeek saved us”  The night of December 24, 2024, was Christmas Eve, though for most people in China it was not a holiday that mattered much. For Julian, it became one of the darkest nights of her life. She was 26, had graduated from Peking University only two years earlier, and had no prior industry experience. Yet she was already one of the earliest employees at Kimi. That night, this very young yet already “senior” employee sat at the long table in a conference room called Radiohead, crying in front of more than 30 colleagues. She still had not delivered a holiday marketing plan that met the standards of the co-founders. Chinese New Year was only a month away. The latest plan had already been revised six times, and now it needed to be upgraded again, perhaps even scrapped entirely. The odds of rebuilding it from scratch and then coordinating product and engineering to execute it in time were slim. But the company had high hopes for growth during the 2025 Lunar New Year period. That mattered because the previous Lunar New Year had been a breakthrough moment for Kimi. It had gone viral in China thanks to its branding around handling “2 million Chinese characters of long-context input,” which was unusually advanced at the time. Consumer users surged, and in the Chinese stock market people even started talking about “Kimi concept stocks,” meaning public companies loosely associated with the trend. That weekly meeting was long and brutal. Around 20 young employees, most as inexperienced as Julian, took turns reporting on everything: social media ads, user operations, PR in China, overseas marketing, all the details. The group discussed everything collectively, and the co-founders made the final calls. Kimi at that point felt like an adolescent: talented, full of potential, but not yet fully in control of itself. Even with a monthly advertising budget of tens of millions of RMB, it still looked clumsy in the face of fast-rising competitors. The meeting ended around 4 a.m. No one knows whether Julian’s final plan would have succeeded. A month later, it no longer mattered. That was when the world first heard the name DeepSeek. Hayley, who worked on growth, went home to Wenzhou for the holiday and found that relatives and friends all asked the same question: “Have you heard of DeepSeek?” It was as if Kimi had suddenly become yesterday’s news. She says that was the hardest Lunar New Year of her life. The silence inside the company was deafening. The annual company meeting is usually held in March, after the holiday. Employees are allowed to challenge management directly. That year, almost every question revolved around DeepSeek. The sharpest question came from the HR team. With complete sincerity, they said the uncomfortable thing out loud: “How are we supposed to answer candidates when they ask: DeepSeek also gave me an offer. Why should I join Kimi instead?” But not everyone reacted the same way. Alex from the algorithm team says that if he felt any strong emotion during the “DeepSeek moment,” it was not fear. It was excitement. That feeling was not just personal. It reflected the mood of much of the algorithm team. DeepSeek had shown that there might be another way: lower-cost strategies, open-source approaches, and a truth many people had doubted before. A little-known Chinese startup, if its technology was strong enough and its model was good enough, could still earn global respect. The product team was not especially anxious either. Kevin, one of the earliest product employees, believed that DeepSeek had broken out because of its model. Once Kimi’s own model capabilities caught up, he believed the product team would have even more room to build useful features on top. No outsider knows exactly what discussions the co-founders had. But the company moved quickly. It adjusted strategy, narrowed focus, and reached something close to full internal alignment. Ask almost anyone inside the company what matters most now, and they will answer without hesitation: the model. From then on, you could feel a growing respect for DeepSeek inside Kimi. Part of it was professional admiration. Part of it was something else. As Alex put it: “In a way, DeepSeek saved us.” ## Taste is all you need “Why are you wearing shoes like that?” After Ezra asked me that, I was more surprised than she was. On her floor of the office, almost everyone keeps a pair of slippers under the desk. Comfortable clothes and shoes, people believe, make you more relaxed, more focused, and more creative. This is the dress code of smart people. I have met many high-achieving students in my life. But the “good students” here are a very different species. When Ezra was in elementary school, she tried to hack the family computer because her parents would not tell her the password. In middle school she became interested in Bitcoin, when one coin cost only a few hundred RMB. She asked her mother for spending money to invest; her mother told her it was a scam. In high school, the first time she ever took a taxi, she sketched out a ride-hailing product concept. Had today’s AI tools existed back then, she says, maybe she could have launched it. Once she finally had some money of her own in college, she put it into the Chinese stock market and lost 90%. That disaster taught her something about the limits of human judgment, and pushed her toward AI. Her view of AGI, or artificial general intelligence, is simple: create “N Einsteins” and use them to solve humanity’s hardest problems. From that point on, she became determined to find a company that would truly push the limits of AGI. This was despite the fact that she had already made her investment losses back in the stock market. Because of her strong academic background, she received offers from many companies. She chose Kimi for one reason: during the interview, she was deeply impressed by founder Yang Zhilin’s understanding of technology and his seriousness about details. She felt he genuinely cared about models. He did not have the restlessness often seen in smart people, nor the utilitarian instinct common in businesspeople. In fact, by the end of the interview, she still did not know he was the founder. Karen’s personality is different but leads to a similar place. He was rebellious from childhood. He argued with teachers. He never listened to his parents. As a student, he insisted on going abroad. After graduating, he insisted on starting a business. The comfortable and stable life offered by a big Chinese tech company made him despair. He did not want a life whose ending was visible from the beginning. I asked him: if given the choice between a guaranteed 60 out of 100, and a 1% chance at 100 out of 100, which would you choose? He chose the latter without hesitation. It was not that he could not tolerate a score of 60. He just hated the certainty of that 100% path. That founder-like DNA forms part of the company’s underlying texture. By rough internal count, at least 50 people at Moonshot AI have founded or joined startups before. Kimi, apparently, likes hiring CEOs. A more accurate way to put it is this: the company shelters a rotating population of gifted drifters. A genius is not necessarily a top student or model employee. What matters is that in some dimension, they can see through time. At a company where around 80% of employees come from China’s elite “985” and “211” universities, Yannis’s résumé does not look especially impressive. Yet as early as 2023, he had already predicted in engineering communities that both DeepSeek and Kimi would rise, at a time when model companies barely had products at all. Another employee, himself born after 2000, noticed Yannis’s insight and recommended him internally. Karen says too many smart people get trapped by systems. First the family, then school, then the workplace. They obey group expectations without realizing it and lose sight of what they actually want. Only a small number try to escape, and even they often go unseen. One of Kimi’s missions, he says, is to see them. Without that instinct, a 17-year-old high school student would never have been brought in as a Kimi intern, collaborated with the team, and published a paper that later drew praise from Elon Musk. The person who put that student’s name first on the paper was Bob, the mentor who first spotted him. There is only a thin line between genius and madness. When an “ununderstood madman” arrives at Moonshot AI, he may suddenly become a world-changing genius. Or perhaps some still-hidden genius can only truly bloom in a place like this. Bob told me that, to some extent, having a big ego is not a problem. It may even be a good sign. If that ego functions as inner drive, if someone believes they must be part of a great mission, that may be exactly the sort of person the company cannot afford to miss. Geniuses are obsessive. Inside this team, training a top AI model is jokingly called “alchemy,” a common Chinese tech term for the mysterious, half-scientific, half-artistic process of model training. But in practice, alchemy means constantly fixing bugs. Once a flagship training run begins, Bob and his teammates fall into the same ritual. The first thing they do every morning is refresh the company’s massive set of internal monitoring dashboards. Hundreds of thousands of metrics. If even one curve spikes abnormally, alarms go off in their heads. Was there a problem in optimization? A flaw in the architecture? A mismatch in numerical precision? They react with almost animal sensitivity. Some people even inspect training data token by token, printing out those that produced extreme gradients and interrogating them like suspects: why did you jump so violently? Everyone who has ever truly participated in “delivering” one of these models has lived through this kind of sleepless tension. It is not really anxiety. It is curiosity driving obsession. That obsessive vigilance is part of what pushed the model toward top-tier performance. Geniuses cluster. Over the past year, more than 100 of Kimi’s hires came through referrals, friends or friends of friends. Inside the company, this is jokingly called “human-to-human transmission.” Trust, because of these dense networks, becomes a natural organizational asset. In essence, Kimi shifts the hardest part of management onto recruiting. If people are brought in by trusted peers, they are more likely to share the same instincts. This is why one word comes up over and over inside the company: Taste. One night in September 2025, several engineers casually launched a small internal project and named it Ensoul. They wanted code sleeping inside files to “come alive” and become a conversational assistant inside the command line. This sensitivity to naming is not accidental. They once had a framework called YAMAHA, short for “Yet Another Moonshot Agent.” Their deepest infrastructure layer was called Kosong, which means “empty” in Malay, inspired by the Buddhist phrase “emptiness is form.” It was meant to suggest a blank sheet of paper with no pre-assigned function, but infinite potential. Taste, in other words, shapes the product itself. While many other companies were shoving chat windows into the command line, Kimi’s engineers thought that was ugly. Real programmers open a terminal to issue commands, not to chat. So Kimi CLI was designed to feel more like a smart shell than a chat interface. It understands commands, but does not force itself into the shape of a conversation box. This minimalism is visible in the code too. The core logic is only about 400 lines of Python, stripped of all unnecessary ornament. The modules are cleanly decoupled. Users can customize functions themselves, or take Kimi apart and reassemble it into their own applications. Even Kimi Agent was once internally associated with the phrase OK Computer, a Radiohead reference, though that name was later changed because it was too obscure for wider adoption. The people who chose names like that did not seem especially interested in maximizing internet traffic. They obeyed their own musical taste and linguistic standards instead. Someone joked that if you measured AI companies by the share of employees who play musical instruments, Kimi might rank first. Taste has become the highest hiring standard, and also the hardest to define. It cannot be quantified, but it is everywhere. ## Generalize, then evolve You may never fully understand what each person at Kimi actually does. The company likes using the word “team” instead of department. At a high level, the main areas are clear enough: algorithms, product and engineering, growth, strategy, operations. But once you try to zoom in and map actual departments or fixed responsibilities, things start to blur. That is because this is an organization with no formal departments, no hierarchy, no titles, no OKRs, and no KPIs. Reporting lines are so simple that they feel almost unreal. For Brandon, this made no sense at all. He had studied at Tsinghua, held management roles at Silicon Valley giants and major Chinese tech firms, and helped build a startup worth around $1 billion. He had spent years in the industry and excelled at technical management. He had led teams of nearly 1,000 people. He hoped to enter AI and apply that experience at scale. Instead, co-founder Zhang Yutong told him that the company did not work that way. The number of people he would likely manage, if he joined, was about two. Still, something about the future pulled him in, and he wanted one more conversation. So in January 2025, during a period of internal doubt and unrest, Brandon met Yang Zhilin, his younger schoolmate from Tsinghua. At the time, Brandon had no idea that Yang’s name would eventually be mentioned in media stories alongside Elon Musk and Jensen Huang. What he remembers most is the very first sentence Yang said after basic greetings: “Reinforcement learning is the future.” The rest of the conversation felt almost like Yang thinking out loud. He was so immersed in his own line of thought that Brandon could not understand much of what he was saying, even though it was all in Chinese. But one thing was unmistakable: for the first time, Brandon felt the knowledge structure and mental models he had built over the past 20 years starting to collapse. Along with them went his ego. When I asked why he eventually joined, he replied in a slightly mysterious tone: Yang Zhilin might become a great prophet, because he is both far-sighted and pure. Later, when the company hesitated because it did not really know how to define his role in such a title-light system, Brandon replied firmly: “Even if you make me clean toilets, I’ll come. And I’ll clean them better than anyone.” Not every former big-tech manager or expert thrives in this environment. Phoebe, born after 2000, moved from the growth team into product and engineering. She describes herself jokingly as “a clueless little girl,” but says something important: in this company, deep experience and strong credentials can actually become a burden. AI is too new. The field is changing too fast. A highly experienced expert may not learn and adapt as fast as a younger person with fewer assumptions. She has seen at least three mid-level or senior big-tech hires fail to “land” after joining. One eventually chose to leave the industry altogether, saying the people around him were just too young and too smart. After being repeatedly outperformed, he gave up. This, he decided, was no longer his era or his industry. After the DeepSeek shock, Phoebe also felt a deep sense of crisis. She decided to abandon ad-buying work and instead try to help the company through product and engineering. She began an intense period of self-study, even streaming herself learning on Bilibili for hundreds of hours. What surprised her most was that the company, from the start, gave her the chance to switch roles without much hesitation. In fact, among the thirty employees I interviewed, more than half had changed responsibilities multiple times. Compared with their previous jobs, perhaps 80% were now doing something completely different. Kimi likes people with generalization ability. In AI, generalization means a model can perform well in new scenarios beyond its training data. It has not merely memorized answers; it has learned underlying structures. The company applies this idea to people too. Mid-level and senior employees from giant firms may have spent too long optimizing for a particular KPI system, a particular reporting language, a particular internal political game. Their “algorithm” becomes overfit to one local optimum. When the environment changes completely, they may fail to adapt. If traditional big-tech workers are like specialized models, then the people Moonshot AI wants are more like base models. First they learn basic rules through supervised fine-tuning. Then, through reinforcement learning and repeated self-play across many tasks, they acquire the ability to transfer across domains. James, a returnee from Silicon Valley, is 26 and says his dream is “to give money to young people.” As a devout believer in AI, he sees his own body as little more than a sensor for an agent to collect information. When playing League of Legends with friends, he records voice and collects physiological data like heart rate and pulse, then analyzes which teammate’s comments affected his emotional state and game performance. His views are so sharp they verge on extreme. He says: if a person starts learning a truly new language after age 14, they will never master it at a native level. AI, he argues, works similarly. Dan, who joined the company right after graduation, says that for the first time in his life he felt true knowledge anxiety. At school, he had only ever worked on “toy models,” around 7 billion parameters, which could be trained in a few days on 32 GPUs. Now he was handling enormous Mixture-of-Experts models with tens of billions of parameters and training datasets measured in trillions of tokens. It felt like jumping straight from a small pond into the Pacific Ocean. To keep up, he threw himself into near self-abusive study. His schedule collapsed. Beijing daytime became Silicon Valley nighttime, then reversed. He stared at training dashboards for hundreds of hours, like a stock trader watching markets with no room to blink. The real challenge was not just workload. He had to do three jobs at once. He had to be an algorithm architect, designing the best plan through a maze of model choices. He had to be a systems engineer, debugging distributed computing problems like a mechanic repairing a pipeline stretched across the globe. He had to be a data curator, performing “alchemy” on giant datasets so the model would score well on benchmarks while also feeling natural and soft in actual conversation. Sometimes that meant emergency surgery mid-training. At one point, key parameters stored in bf16 precision started behaving dangerously. The team made a snap decision to switch to fp32 precision halfway through training, just to stabilize the run. Dan says that if all you can do is write algorithms, or build systems, or clean data, you will never produce a top model. There is no excuse here of “I only handle this part.” The company expects you to integrate algorithm, engineering, and data work across multiple worlds. It is like doing several jobs at once. But that kind of intense cross-training can give you years’ worth of growth in a very short time. So anyone trying to join Kimi faces a brutal test. There are no OKRs, no KPIs, no office politics, no manipulative managers, not even clock-in attendance. But if you are not AI-native, if you cannot generalize, if you cannot continuously reinforce and adapt, then you may struggle to find meaning for your existence here. ## “There’s no bureaucrat smell here” Most brands want a story. But nearly every Kimi employee gently warned me: don’t write about Pink Floyd, or the piano near the office entrance. Their view is that people who get it, get it. People who don’t, don’t need to. The names Moonshot and Kimi have nothing directly to do with AI or technology. But if the company talked too much about its connection to rock music or art, it would start to feel self-conscious and pretentious. Better, they seem to think, to be beautiful without trying to explain the beauty. Win, another post-2000s employee who had escaped from a giant tech company, told me this place is bizarre because people can actually get work done without endless meetings. At his former employer, daytime was for meetings and nighttime was for work. He learned a simple lesson: if your energy goes mainly into coordinating relationships around production, there is very little room left to improve actual productivity. This is part of what an AI-native organization looks like. More than ten employees told me explicitly that they increasingly prefer dealing with AI over dealing with humans. AI feels more reliable and simpler. That tendency also fits the company’s broader introverted character. One person used a gentler word: shy. In group chats, everyone can be lively and expressive. In person, many are quiet. Kimi does not organize many cultural activities. Aside from the annual meeting, the most recent group event had simply been massages in the office. Introversion does not mean a lack of communication or energy. Even though no one was required to talk to me, not a single person said no. In group chats, information flies constantly, along with all kinds of abstract emoji. No one’s messages are left hanging in silence. And if you need help from someone else to get work done, the process is simple: ask them directly. No need to go through a manager. No need for approval. No need for a coordination meeting. No need to break through departmental walls. Kimi has no departmental walls. In some sense, it does not even have departments. Yang Zhilin’s status message is just four words: Communicate directly. Still, everyone acknowledges that the company has changed continuously since its founding. Some changes were proactive, some reactive, and some even seemed like reversals. The company moved from heavy ad spending to model focus, from insisting on closed source to embracing open source, from chatbot products to Kimi Agent, Kimi Code, and Kimi Claw, from consumer to enterprise and back again. Not every shift stands up perfectly to scrutiny. Yet in Ezra’s mind, one thing has remained constant: respect for facts. All those changes, she believes, had only one cause and one purpose: to make the company align better with objective reality. The company tolerates ego, but it does not like hiring people who place themselves above facts. From the co-founders down, people are relatively easy to persuade, as long as the facts are clear enough. That willingness, employees say, comes from an intense commitment to truth, reality, and what is real. Truly smart people are not wounded by honest feedback. Another condition for this level of honesty is that the company has no horse-race system, no zero-sum competition, no major internal conflicts of interest. People willingly share research findings and technical detail without expecting payment or credit. Early on the company had its own community; today it still promotes a community culture. Shared information and shared knowledge speed up everyone’s learning, which in the end benefits everyone. Win says toxic culture is contagious. Good culture is contagious too. Someone used the word “solidarity” to describe the atmosphere, a word that sounds almost old-fashioned when applied to a startup. But the company operates in a harsh environment. Outside are giant competitors. Inside are the pressures of being squeezed by established tech firms. Compute resources are limited. Those constraints, if anything, seem to increase cohesion. At the root of it all, people are the only truly important asset in an organization. Recently, Florence was approached by a competing company offering double her salary. She rejected it immediately. Her reason was simple: “There’s no ‘officialdom smell’ here.” That phrase is hard to translate directly. In Chinese internet slang, it refers to the stale, hierarchical, self-important atmosphere associated with bureaucracy, performative authority, and status games.  The company's new office. ## “I don’t know how she endured it” At the beginning of this reporting process, I was extremely nervous. I was about to interview some of the smartest AI people in the world. I am a humanities person. I have never worked in tech. My knowledge of AI is limited. But when I actually started talking with young experts from the algorithm and product-engineering teams, I realized they were the ones who seemed nervous. They were afraid I would feel awkward if I did not understand their terminology. So first they would translate English into Chinese, and then translate that Chinese into a second, even simpler Chinese I could understand. That instinct to protect was moving. Before I started the interviews, the company gave me only one instruction: protect everyone. So I tried to avoid questions that were too sensitive or likely to hurt people. Even so, Ty, during a phone interview, could not fully hide a small emotional tremor. When he first joined the company and was going through the difficult onboarding process, he struggled badly. At one point he felt he could not continue and even thought about resigning. Then one week, at the company meeting, he watched Annie, a woman who had graduated only two years earlier, finally push a difficult project forward after countless setbacks and internal doubts. Seeing that, he felt he could not give up either. He was older than she was, had more life experience, yet in terms of sheer stamina and willpower, he felt weaker. He said: “I don’t know how she endured it.” In fact, Ty was not the only one who had thought about leaving. Annie had too. For a long time, she was trying to build a business line overseas from zero to one and made no real breakthrough. To make things worse, colleagues from other teams, with good intentions, directly told her to abandon what they viewed as a meaningless effort. She says she cried more at Kimi than for any other company, or for any ex-boyfriend she had ever had. It was not as though she lacked alternatives. She already had a better-paying offer elsewhere. But she says she simply could not persuade herself to go work for someone else. She wanted one more conversation with Zhang Yutong. Afterward, she decided to stay. She did not tell me what was said in that conversation. She only said: Yutong is the strongest boss I have ever seen, the fastest at iterating, with the highest ceiling. Following her is how I can raise my own ceiling. Then Annie repeated the same line: “I don’t know how she endured it.” Once you gather enough material, you notice certain sentences recurring. And the most repeated phrases often reveal the deepest common qualities of a team. Bob, who had been pulled back to China by Yang Zhilin and gave up the chance to pursue a PhD in the United States, joined the company on day one. If anyone understands the company deeply, he does. When I asked him the same question I asked everyone else, what is the team’s most important quality, he thought for about two minutes and answered with one word: Resilience. For a company only three years old, talking about resilience may sound like a luxury. But he means it sincerely. Smart and brave, he says, are sometimes opposites. The smarter you are, the more clearly you see the risks, and the easier it becomes to walk away. Foolish persistence will not succeed either. So only those who see the truth, calculate the odds of failure, and still continue deserve to be called resilient. Inside the company, there is a story known as “three trips to the cliff of reflection.” In May 2023, Freddie and his colleagues were given a task that seemed impossible: make AI read and understand 128K context in a single pass, meaning hundreds of book pages, at a time when the industry standard was closer to 4K. He quickly designed a solution called MoBA v0.5, but it required rewriting the underlying training framework while the main model was already halfway through training. The cost was too high, so the idea was shelved. That was the first trip to the “cliff of reflection.” Half a year later he returned with version 1, now designed to continue training from the existing model. It worked on small models, but when tested on the large one it hit a loss spike and kept failing. The project was forced back to the cliff a second time, for another six months. It even missed the company’s 200,000-character product milestone. But the team was not disbanded. Instead, the company launched what it called a “saturation rescue,” gathering technical experts from everywhere to attack the problem together. They rewrote core logic and finally got version 2 to pass the classic long-context “needle in a haystack” test. Just when launch seemed close, a third blow arrived. During supervised fine-tuning, the model performed poorly on long-summary tasks because the training signals were too sparse. By then huge resources had already been invested. Still, the engineers went back to the cliff again, searched for a solution, and eventually fixed the issue by changing the attention mechanism in the final layers. Three retreats. Three returns. At the end of the interview, I asked Freddie the ultimate question: how would you describe this company? He answered in two words: Moon landing. Why moon landing? He quoted the famous line from John F. Kennedy: > We choose to go to the moon in this decade and do the other things, not because they are easy, but because they are hard.  All the company meeting rooms are named after musical acts. ## Genius Swarm In the end, I did not disturb or attempt to probe the co-founders themselves. Externally, they remain almost invisible. They dislike interviews and have no interest in personal fame. Internally, though, they are everywhere. In an extremely flat organization, you need superbrains at the center. Otherwise vitality turns into chaos. Because there is little middle management, each co-founder interfaces directly with around 40 to 50 employees and stays close to both the technical and business front lines. That is how the company keeps decision-making and execution aligned. All five co-founders came from Tsinghua University. But biological limits still exist. Human attention spans are finite. Management range is finite. Once the company reached a RMB 120 billion valuation and grew past 300 people, even these superbrains began to strain under the load. And it is not just the founders. This is an infinite game driven by self-motivation. If every member is effectively carrying RMB 400 million of valuation, then each person is expected to create an extraordinary amount of value. The revolutionary variable is the toolset. Kimi does not actually run on extreme working hours. Employees are allowed to wake naturally. They are not required to stay in the office until dawn every night. Leo from the product team says he commands “an army” now, meaning AI agents. Imagine this scenario: Leo wakes up at 10 a.m. and walks into the office. His task is to analyze user feedback from five global markets over the past 24 hours and decide this week’s product priorities. In the past, that would have taken three people two days. Now he launches three agents. A strategy agent scans 3,000 feedback items and filters for high-priority requests related to long-context interruption. A translation agent interprets Japanese dialect and Korean honorifics in real time and marks true emotional intensity. A competitor agent monitors updates from Cursor and ChatGPT and produces a technical comparison. Leo does only three things himself. He rejects one sarcastic comment that the system had misread as sincere. He flags a screenshot containing an unreleased UI. He confirms the top three needs recommended by the agents. By 11:30 a.m., the product requirements document is already finished. Meanwhile, a coding agent has generated about 70% of the base implementation, leaving only the more creative design work for afternoon discussion with human engineers. Humans set the rules. Silicon-based systems execute them. The organization becomes a container for algorithms. In an AI-native company, using agents skillfully and embedding them deeply into workflows is not optional. It is part of the job. The model is not only the goal. It is also the tool. Whether by directly improving productivity or by fundamentally changing management structure, AI’s logic has already entered the bones of this company. Just as the company builds an Agent Swarm, the team itself begins to resemble a Genius Swarm: many independent geniuses working in parallel, coordinating seamlessly. Still, such a flat structure has built-in fragility. When I asked whether this model would remain sustainable if the company grew from 300 people to 3,000, most people answered cautiously. History is not encouraging. Similar experiments in extreme flatness, like holacracy or Haier’s internal contract-cell structures, often hit decision bottlenecks once they pass around 500 people. When there are too many information nodes, “direct communication” starts turning into information overload. A more immediate pain point is the personal experience of weightlessness. Without hierarchy to buffer uncertainty, confusion about direction is felt directly by each individual. One former employee who eventually returned to big tech put it bluntly: without top-down OKRs and KPIs, some mornings you walk into the office not knowing what you should do. No one necessarily tells you whether you are doing well. That lack of feedback creates insecurity. It can make people nostalgic for the clear reporting lines, review points, and measurable outputs of giant tech companies. Those cumbersome structures, after all, do provide one essential thing: a baseline of certainty. Where is the goal? What counts as completion? How will performance be judged? In a large firm, all that is visible. That is not Stockholm syndrome, the person said. It is basic organizational physics. If Alibaba is like a finely calibrated promotion conveyor belt, ByteDance like a ruthless battle corps with strong objectives, and Tencent like a more forgiving professional academy, then Moonshot AI is like a primeval forest. Geniuses may find a hunting path. Ordinary people may just wander in the fog. ## The necessary “two-dimensional foil” No departments. No titles. No evaluations. The AI-native organizational model is anti-bureaucratic and intentionally unstructured. Large companies can no longer pivot toward it easily. Small companies often miss the window because they expand into traditional structures too quickly. This is an asymmetric war. Here the author turns to a famous science-fiction reference from The Three-Body Problem. In that story, an advanced civilization casually uses a weapon called a two-dimensional foil, which collapses the solar system from three dimensions into two. Planets, stars, and humans all become a flat image without thickness. Moonshot AI, the author argues, is deliberately throwing such a “two-dimensional foil” at itself. Not to destroy an opponent, but to flatten the organization in pursuit of maximum efficiency. No vertical depth of hierarchy. No horizontal walls of departments. No three-dimensional tangles of office politics. Only “model” and “intelligence” facing each other directly in the simplest possible form. In the age of AI, every startup, the author argues, is being forced to throw such a foil at itself. The rise of one-person companies reflects the same generational explosion of AI-native talent. If technology can compress organizational capability into the individual, then many of the middle layers of management simply evaporate. The organization gets flattened. There is no depth left for detours. Everyone is forced to face the problem itself. That may be the hard rule governing the evolution of organizations in the business world. Everyone, eventually, will be folded. Once people are exposed on the same plane, one person radiating influence over fifty others no longer looks like a managerial miracle. It becomes normal. The distance from center to edge is redefined. People who depend on titles and OKRs as coordinates may suffocate instantly. But geniuses, on this exposed flat surface, can violently dismantle intelligence itself, while the “guardians” clear away noise and entropy, seeing themselves, not without humility, as pioneers widening the boundary of human civilization. And yet the transition from three dimensions to two cannot be reversed. That means Kimi cannot go backward. Every strategic adjustment becomes a chaotic iteration with high stakes. Competitors can still turn slowly inside a maze. But if Moonshot AI tries to expand recklessly in size, it may tear itself apart structurally. This act of self-flattening is only acceptable because it is in service of something more radical. The endpoint of lowering the organization’s dimension is raising the dimension of intelligence. Only if model intelligence crosses the critical threshold, rising high enough to escape the gravity well of all carbon-based organizations, can Moonshot AI truly crush the organizational advantages of its competitors and justify this irreversible gamble. At that point, debates over management span or org charts no longer matter. It would be like asking what dimension the Three-Body Problem civilization inhabits, when the real point is that its dimensional weapon has already rewritten the rules of war. Then “Moonshot AI” would stop being a metaphor. It would become a higher-dimensional light source, illuminating the dark side of the intelligence universe. All the organizational pain that came before would be no more than the heat shield burning off as the lunar module passed through the atmosphere. Either they become godlike through ascent. Or they are sealed away in collapse. There is no third path. All of the English names used are pseudonyms. ## 相关链接 - [Rui Ma](https://x.com/ruima) - [@ruima](https://x.com/ruima) - [589K](https://x.com/ruima/status/2039245985520681257/analytics) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [3:38 PM · Apr 1, 2026](https://x.com/ruima/status/2039245985520681257) - [589.4K Views](https://x.com/ruima/status/2039245985520681257/analytics) - [View quotes](https://x.com/ruima/status/2039245985520681257/quotes) --- *导出时间: 2026/7/22 09:40:16* --- ## 中文翻译 # Kimi 内部100小时 **作者**： Rui Ma **日期**： 2026-04-01T07:38:24.000Z **来源**： [https://x.com/ruima/status/2039245985520681257](https://x.com/ruima/status/2039245985520681257) ---  本文由中文翻译而来，原载于《人物》，这是一家以报道特稿和人物专访著称、备受尊敬的中国杂志。文章已为可能不熟悉中国背景、公司文化或相关典故的读者做了改写。我使其更易于阅读，在必要之处添加了简要背景说明，并润色了一些直译会显得生硬的表述。 文：刘谟 编：金匝 原文于2026年3月31日由《人物》以中文首次发表 2026年的春天对Kimi格外眷顾。 短短几个月里，Kimi背后的公司似乎接连跨过一个又一个里程碑。它的营收、融资和估值不断刷新纪录。一篇由17岁高中生实习生参与合著的研究论文，获得了包括埃隆·马斯克在内的硅谷人士的赞誉。而估值约500亿美元的美国编程创业公司Cursor，则被中国观察者指责其本质上是在“套壳”或严重依赖Kimi的模型作为自身产品体验的一部分。换言之，Kimi似乎突然同时在三条战线上告捷：资本、技术和商业落地。 这家创业公司只有三岁。它的估值已超过1200亿元人民币，约合160亿美元。在全球AI的故事里，它正变得不容忽视。 然而，Kimi背后的月之暗面（Moonshot AI），依然充满神秘感。 我获准花100个小时从内部观察这家公司。作为一名独立撰稿人，我被允许采访任何愿意交谈的员工，旁听任何不涉及商业机密的会议，并在事后自由写作。没有人会修改我的稿子，我也不会获得报酬。事实证明，这非常符合这家公司的作风。 走进办公室，感觉就像站在风暴的中心。 在中心处，一切出奇地平静。办公桌一片安静，只有零星的键盘声打破寂静。偶尔能听到有人笑。而外界的喧嚣——流言、争论、炒作、模仿和无休止的评论——似乎在这里不留一丝痕迹。 员工只有300多人，平均年龄不到30岁。如果用公司估值除以人数，每个人肩上实际上扛着将近4亿元人民币的企业价值。 约80%的员工是中文互联网俚语所说的"I人"，借用MBTI的说法，即内向者。人们并肩而坐，但比起说话，他们更习惯打字。在这里，内向不被视为缺陷，它几乎是一种运行协议。 我不禁回想2024年我第一次到访的那个夜晚，那时风暴才刚刚开始聚集。当时，我并没有留下什么特别好的第一印象。 ## “DeepSeek救了我们”  2024年12月24日的那个夜晚是平安夜，尽管对大多数中国人来说，这并不是一个很重要的节日。而对Julian来说，那成了她人生中最黑暗的夜晚之一。 她26岁，两年前才从北京大学毕业，没有任何行业经验，却已经是Kimi最早的员工之一。那晚，这位如此年轻却已是“资深”员工，坐在一间名为Radiohead的会议室的长桌旁，在30多位同事面前哭了。 她仍然没能拿出一份符合联合创始人标准的节日营销方案。 离春节只有一个月了。最新的方案已经改了六遍，现在又要再改，甚至可能推倒重来。从零重建方案、再协调产品和工程团队及时执行，胜算渺茫。但公司对2025年春节期间的增长寄予厚望。 这很重要，因为上一个春节正是Kimi的突破时刻。它凭借“200万字长文本输入”这一在当时极为领先的卖点在中国爆红。消费级用户激增，A股市场上甚至开始出现“Kimi概念股”的说法，指那些与这股热潮沾边的上市公司。 那场周会漫长而残酷。 大约20名年轻员工——大多和Julian一样缺乏经验——轮流汇报所有事项：社交媒体广告、用户运营、国内公关、海外营销，事无巨细。所有人集体讨论一切，联合创始人做最终决定。 那时的Kimi像一位少年：才华横溢、潜力无限，却还未完全掌控自己。即便每月有数千万元人民币的广告预算，面对快速崛起的竞争对手，它仍显得笨拙。 会议在凌晨4点左右结束。 没有人知道Julian的最终方案能否成功。一个月后，这已无关紧要。 因为就在那时，世界第一次听到了DeepSeek的名字。 负责增长的Hayley回温州老家过年，发现亲戚朋友问的都是同一个问题：“你听说过DeepSeek吗？”仿佛Kimi一夜之间成了昨日新闻。 她说那是她人生中最难熬的一个春节。公司内部的沉默震耳欲聋。 公司年会通常在春节后的三月举行，员工可以直接向管理层提出质疑。那一年，几乎每一个问题都围绕着DeepSeek。 最尖锐的问题来自人力资源团队。他们以完全的诚意，把那个令人难受的问题说了出来： “当候选人问我们：DeepSeek也给我发了offer，我为什么要加入Kimi而不是他们？我们该怎么回答？” 但并非每个人的反应都相同。 算法团队的 Alex 说，如果他在“DeepSeek 时刻”感受到任何强烈的情绪，那不是恐惧，而是兴奋。 这种感觉不仅是个人的。它反映了算法团队许多人的心境。DeepSeek 已经证明，或许存在另一条路：更低成本的策略、开源的方式，以及一个许多人此前曾怀疑的真理。一家鲜为人知的中国初创公司，只要技术足够强、模型足够好，依然可以赢得全球的尊重。 产品团队也没有特别焦虑。作为最早的一批产品员工之一，Kevin 认为 DeepSeek 的爆发靠的是它的模型。一旦 Kimi 自身的模型能力追上来，他相信产品团队将有更大的空间去在其之上构建有用的功能。 没有任何外部人士确切知道联合创始人们讨论了什么。但公司行动迅速。它调整了战略，收窄了焦点，并达成了近乎完全的内部一致。 几乎问公司里的任何一个人，现在最重要的是什么，他们都会毫不犹豫地回答：模型。 从那时起，你能感觉到 Kimi 内部对 DeepSeek 与日俱增的敬意。一部分是专业上的钦佩。另一部分则是别的东西。 正如 Alex 所说： “某种意义上，DeepSeek 救了我们。” ## 品味就是一切 “你怎么穿这种鞋？” Ezra 问出这句话后，我自己都比她还惊讶。在她所在的办公楼层，几乎每个人都在桌子底下放一双拖鞋。人们相信，舒适的衣服和鞋子能让你更放松、更专注、更有创造力。 这就是聪明人的着装规范。 我这一生见过许多成绩优异的学生。但这里的“好学生”是完全不同的物种。 Ezra 上小学时，因为父母不肯告诉她密码，她试图破解家里的电脑。初中时她对比特币产生了兴趣，那时一枚币只要几百元人民币。她向母亲要零花钱去投资；母亲说那是骗局。高中时，她第一次坐出租车，就勾勒出了一个网约车产品的概念。她说，如果当时有今天这样的 AI 工具，也许她真能把它做出来。等到大学终于有了自己的钱，她把钱投进了中国股市，亏掉了 90%。 那次惨败让她对人类判断力的局限有了认识，并把她推向了 AI。 她对 AGI（通用人工智能）的看法很简单：创造出“N 个爱因斯坦”，用他们去解决人类最难的问题。从那一刻起，她下定决心要找一家真正推动 AGI 边界的公司——尽管她当时已经在股市里把亏损的钱赚了回来。 凭借出色的学术背景，她收到了许多公司的 offer。她选择 Kimi 只有一个原因：面试中，创始人杨植麟对技术的理解和对待细节的认真态度给她留下了深刻印象。她感到他是真心在意模型的。他没有聪明人身上常见的那种浮躁，也没有生意人身上常见的功利本能。事实上，直到面试结束，她都还不知道他就是创始人。 Karen 的性格不同，但殊途同归。 他从小叛逆。他和老师争论。他从不听父母的话。学生时代，他坚持要出国。毕业后，他坚持要创业。中国大厂提供的那种舒适稳定的生活让他感到绝望。他不想要一种从一开始就能看到结局的人生。 我问他：如果在稳拿 60 分和只有 1% 概率拿到 100 分之间做选择，你会选哪个？ 他毫不犹豫地选择了后者。 并不是他无法容忍 60 分。他只是厌恶那条 100% 确定路径的确定性。 那种创始人般的 DNA 构成了公司底色的一部分。据粗略的内部统计，月之暗面至少有 50 人曾经创立或加入过创业公司。 显然，Kimi 很喜欢招 CEO。 更准确的说法是：这家公司收容着一群不断流动的天才流浪者。天才未必是好学生或模范员工。重要的是，在某个维度上，他们能看穿时间。 在一家约 80% 员工都来自中国精英“985”“211”高校的公司里，Yannis 的履历并不算特别亮眼。然而早在 2023 年，他就在工程师社区中预言了 DeepSeek 和 Kimi 的崛起，而那时模型公司几乎连产品都还没有。另一位 00 后员工注意到了 Yannis 的洞察力，并为他做了内部推荐。 Karen 说，太多聪明人被系统困住了。先是家庭，然后是学校，然后是职场。他们在不知不觉中服从群体的期待，迷失了自己真正想要的东西。只有少数人试图逃离，而即便是他们，也常常不被看见。 他说，Kimi 的使命之一，就是看见他们。 没有这种直觉，一个 17 岁的高中生就不会被招进 Kimi 实习、与团队合作，并发表一篇后来获得埃隆·马斯克赞扬的论文。把那个学生的名字放在论文第一作者位置的人，是最早发现他的导师 Bob。 天才与疯狂之间只有一线之隔。当一个“不被理解的疯子”来到月之暗面，他可能会突然成为改变世界的天才。又或许，某些仍被隐藏的天才，只有在这样的地方才能真正绽放。 Bob 告诉我，某种程度上，自我（ego）强大并不是问题。它甚至可能是一个好迹象。如果这种自我成为内在驱动力，如果一个人相信自己必须参与一项伟大的使命，那也许恰恰是公司最不能错过的那种人。 天才都是偏执的。 在这个团队内部，训练顶级 AI 模型被戏称为“炼丹”——这是中国科技圈对模型训练这一神秘过程、半科学半艺术过程的通用说法。但在实践中，炼丹意味着不停地修 bug。 一旦一次旗舰级训练启动，Bob 和他的队友们就会陷入同样的仪式。每天早上他们做的第一件事，就是刷新公司庞大的内部监控仪表盘。数十万个指标。哪怕只有一条曲线异常飙升，警报就会在他们脑中拉响。是优化出了问题？架构有缺陷？数值精度不匹配？ 他们的反应近乎动物般的敏感。 有人甚至会逐个 token 地检查训练数据，把那些产生了极端梯度的 token 打印出来，像审讯嫌疑犯一样质问它们：你为什么跳得这么剧烈？ 每一个真正参与过这些模型“交付”的人，都经历过这种无眠的紧张。那其实不是焦虑，而是好奇心驱动的痴迷。正是这种偏执的警觉，推动模型迈向了顶级性能。 天才都是聚堆的。 过去一年里，Kimi 有超过 100 名员工是通过内推、朋友或朋友的朋友加入的。在公司内部，这被戏称为“人传人”。 信任，由于这些密集的关系网络，成为一种天然的组织资产。 本质上，Kimi 把管理中最难的部分转移到了招聘环节。如果人是被信任的同伴引荐进来的，他们更可能拥有相同的直觉。这就是为什么公司内部有一个词被反复提起： 品味（Taste）。 2025 年 9 月的一个夜晚，几位工程师随手启动了一个小型内部项目，并命名为 Ensoul。他们希望沉睡在文件中的代码能够“活过来”，成为命令行里的对话助手。 这种对命名的敏感并非偶然。 他们曾有一个名为 YAMAHA 的框架，是 "Yet Another Moonshot Agent" 的缩写。他们最深层的底层设施叫做 Kosong，在马来语中意为“空”，灵感来自佛语“色即是空”。它意在暗示一张没有预设功能、却蕴藏无限可能的白纸。 换句话说，品味塑造了产品本身。 当许多其他公司把聊天窗口硬塞进命令行时，Kimi 的工程师觉得那样很丑。真正的程序员打开终端是为了执行命令，而不是聊天。因此 Kimi CLI 的设计更像一个智能 shell，而不是聊天界面。它能理解命令，却不会强迫自己变成一个对话框的形状。 这种极简主义也体现在代码里。核心逻辑只有约 400 行 Python，去除了所有不必要的修饰。模块之间干净地解耦。用户可以自行定制功能，也可以把 Kimi 拆开，重新组装进自己的应用里。 就连 Kimi Agent 内部也曾与 OK Computer 这个短语联系在一起——这是对 Radiohead 乐队的致敬，不过后来因为这个名字太过冷僻、不利于推广而改掉了。选择这类名字的人，似乎对最大化互联网流量并不特别感兴趣。他们遵从的是自己的音乐品味和语言标准。 有人开玩笑说，如果按会演奏乐器的员工比例来衡量 AI 公司，Kimi 可能排第一。 品味已成为最高的招聘标准，也是最难定义的。 它无法被量化，却又无处不在。 ## 先泛化，再进化 你可能永远无法完全弄清 Kimi 的每个人具体在做什么。 这家公司喜欢用“团队”而不是“部门”这个词。从宏观上看，主要领域足够清晰：算法、产品与工程、增长、战略、运营。但一旦你试图深入细究，去描绘实际的部门或固定职责，一切就开始变得模糊。 这是因为这是一个没有正式部门、没有层级、没有头衔、没有 OKR、也没有 KPI 的组织。汇报关系简单到近乎不真实。 对 Brandon 来说，这完全说不通。 他曾在清华求学，在硅谷巨头和中国大型科技公司担任过管理职务，还参与打造过一家估值约 10 亿美元的初创公司。他在行业里浸淫多年，擅长技术管理，带领过近 1000 人的团队。他希望进入 AI 领域，把那些经验大规模地应用起来。 然而，联合创始人张宇桐告诉他，公司不是那样运作的。如果他加入，他大概率管理的人数大约是两个。 尽管如此，某种关于未来的东西吸引了他，他想再谈一次。 于是 2025 年 1 月，在一段公司内部充满疑虑和动荡的时期，Brandon 见到了他的清华学弟杨植麟。 当时，Brandon 完全不知道杨植麟的名字日后会在媒体报道中与埃隆·马斯克和黄仁勋并列出现。他印象最深的，是杨植麟在简单寒暄之后说的第一句话： “强化学习是未来。” 接下来的谈话几乎像是杨植麟在自言自语。他沉浸在自己的思路里，以至于 Brandon 尽管听的全程都是中文，却没能听懂多少。 但有一点是确定无疑的：Brandon 第一次感觉到，自己在过去 20 年里构建的知识结构和思维模型开始崩塌。随之崩塌的，还有他的自负。 当我问起他为什么最终加入时，他略带神秘地回答：杨植麟可能会成为一位伟大的先知，因为他既远见，又纯粹。 后来，当公司因为实在不知道如何在这样一个头衔稀少的体系里定义他的角色而犹豫时，Brandon 坚定地回答： “就算让我打扫厕所，我也会来。而且我会扫得比任何人都好。” 并非每一位前大厂经理或专家都能在这种环境中如鱼得水。 00后出生的Phoebe从增长团队转入了产品和工程。她开玩笑地自称“一个懵懂的小姑娘”，但她道出了一点很重要：在这家公司，深厚的经验和亮眼的资历反而可能成为一种负担。 AI太新了。这个领域变化太快。一位经验极其丰富的专家，学习和适应的速度可能不如一个先入之见更少的年轻人。 她亲眼见过至少三位大厂中层或高级员工入职后没能“落地”。其中一人最终选择彻底离开这个行业，说身边的人实在太年轻、太聪明了。在被反复碾压之后，他放弃了。他断定，这已经不再是他的时代，也不再是他的行业。 DeepSeek带来的震撼之后，Phoebe也感到了深深的危机感。她决定放弃投放工作，转而尝试通过产品和工程来帮助公司。她开始了一段高强度的自学时期，甚至把自己学习的过程在B站上直播了数百个小时。 最让她意外的是，公司从一开始就没怎么犹豫，就给了她转岗的机会。 事实上，在我访谈的三十名员工中，超过一半的人都曾多次变换职责。与之前的工作相比，他们中大约80%的人现在做的事情已完全不同。 Kimi喜欢有泛化能力的人。 在AI领域，泛化指的是模型能够在训练数据之外的新场景中表现良好。它不只是记住了答案，而是学到了底层的结构。 公司把这一理念也用在了人身上。 来自巨头公司的中高级员工，可能在某一套特定的KPI体系、某一套特定的汇报话术、某一场特定的内部政治博弈中优化得太久。他们的“算法”过拟合于某一个局部最优解。当环境彻底改变时，他们可能无法适应。 如果说传统大厂员工像是专用模型，那么月之暗面想要的人更像是基座模型。首先通过监督微调学习基本规则，然后通过强化学习以及在众多任务上反复的自我对弈，获得跨领域迁移的能力。 从硅谷归来的James今年26岁，他说自己的梦想是“给年轻人发钱”。 作为AI的虔诚信徒，他把自己的身体看作是一个智能体收集信息的传感器。和朋友打英雄联盟时，他会录下语音，并采集心率和脉搏等生理数据，然后分析哪位队友的言论影响了他的情绪状态和游戏表现。 他的观点犀利到近乎极端。他说：如果一个人在14岁之后才开始学习一门真正新的语言，就永远无法达到母语水平。他认为，AI的工作原理与此类似。 毕业后就加入公司的Dan说，他有生以来第一次感到了真正的知识焦虑。 在学校时，他只做过“玩具模型”，大约70亿参数，在32块GPU上几天就能训练完。而现在，他要面对的是数百亿参数的巨型混合专家模型，训练数据集以数万亿token计。这感觉就像直接从小池塘跳进了太平洋。 为了跟上节奏，他投入到近乎自虐的学习中。他的作息彻底紊乱。北京的白天变成了硅谷的夜晚，然后又颠倒回来。他盯着训练仪表盘长达数百个小时，像一个没有眨眼余地的股票交易员盯着盘面。 真正的挑战不只是工作量。他必须同时做三份工作。 他得是算法架构师，在模型的迷宫般的选择中设计出最优方案。他得是系统工程师，像修理横跨全球的管道的技工一样调试分布式计算问题。他还得是数据策展人，对庞大数据集进行“炼金术”，让模型在基准测试中取得高分，同时在真实对话中又显得自然、柔和。 有时这还意味着在训练中途进行紧急手术。有一次，以bf16精度存储的关键参数开始出现危险的表现。团队当机立断，在训练中途切换到fp32精度，只为让运行稳定下来。Dan说，如果你只会写算法、只会搭系统、或者只会清洗数据，你永远做不出顶级模型。在这里没有“我只负责这一块”的借口。 公司期望你把算法、工程和数据工作融为一体，横跨多个世界。这就像同时做好几份工作。但这种高强度的交叉训练，能让你在极短的时间内获得相当于数年的成长。 所以，任何想加入Kimi的人都要面对一场残酷的考验。 这里没有OKR，没有KPI，没有办公室政治，没有玩弄权术的管理者，甚至没有打卡考勤。但如果你不是AI原住民，如果你没有泛化能力，如果你不能持续强化和适应，那么你可能很难在这里找到自己存在的意义。 ## “这里没有官僚味” 大多数品牌都想要一个故事。 但几乎每一位Kimi员工都委婉地提醒我：别写平克·弗洛伊德，也别写办公室入口处的那架钢琴。 他们的看法是，懂的人自然懂，不懂的人也不需要懂。Moonshot和Kimi这两个名字与AI或技术没有直接关系。但如果公司过多谈论自己与摇滚乐或艺术的渊源，就会开始显得刻意和做作。他们似乎认为，美就美在不刻意解释这份美。 Win是另一位从大厂逃离出来的00后员工，他告诉我，这个地方很奇特，因为人们不需要无休止的会议就能把事情办成。 在他的前雇主那里，白天用来开会，晚上用来干活。他学到了一个简单的道理：如果你的精力主要花在围绕生产协调人际关系上，那么留给提升实际生产力的空间就所剩无几了。 这就是AI原生组织的一部分样子。 十多名员工明确告诉我，他们越来越倾向于与AI打交道而不是与人打交道。AI让人感觉更可靠、更简单。这种倾向也与公司整体偏内向的性格相符。有人用了一个更温和的词：害羞。 在群聊里，每个人都可以活泼而善于表达。面对面时，很多人则很安静。Kimi不太组织文化活动。除了年会之外，最近一次集体活动只是办公室里的按摩。 内向并不意味着缺乏沟通或缺乏能量。 尽管没有人被要求接受我的采访，但没有一个人说不。在群聊里，信息不断飞驰，还夹杂着各种抽象表情包。没有谁的消息会被晾在那里无人回应。 而如果你的工作需要别人的帮助，流程很简单：直接去问。 不需要经过经理。不需要审批。不需要开协调会。不需要打破部门壁垒。 Kimi没有部门壁垒。在某种意义上，它甚至没有部门。 杨植麟的签名状态只有四个字： 直接沟通。 尽管如此，所有人都承认，公司自成立以来一直在不断变化。 有些变化是主动的，有些是被动的，有些甚至看起来像倒退。公司从重金投放广告转向专注模型，从坚持闭源转向拥抱开源，从聊天机器人产品转向Kimi Agent、Kimi Code和Kimi Claw，从消费端转向企业端再转回来。并非每一次转向都经得起完全的推敲。 然而在Ezra心中，有一件事始终未变：对事实的尊重。 她相信，所有那些变化只有一个原因和一个目的：让公司更契合客观现实。 公司容忍自我，但不喜欢聘用把自己凌驾于事实之上的人。 从联合创始人往下，人们相对容易被说服，只要事实足够清楚。员工们说，这种意愿源于对真相、现实和真实的强烈追求。真正聪明的人不会被诚实的反馈所刺伤。 这种坦诚程度的另一个前提是，公司没有赛马机制，没有零和竞争，没有重大的内部利益冲突。人们心甘情愿地分享研究成果和技术细节，不求回报或不求署名。公司早期有自己的社区，如今仍然倡导社区文化。信息与知识的共享加速了每个人的学习，最终惠及所有人。 Win说，有毒的文化会传染。好的文化也会传染。 有人用“团结”一词来形容这里的氛围，这个词用在一家创业公司身上听起来几乎有些过时。但这家公司所处的环境很严酷。外面是巨头竞争对手，里面是被成熟科技公司挤压的压力。算力资源有限。这些约束似乎反而增强了凝聚力。 归根结底，人才是组织中唯一真正重要的资产。 最近，Florence被一家竞争对手公司以双倍薪资挖角。她当即拒绝了。她的理由很简单： “这里没有‘官味’。” 这个词很难直接翻译。在中国网络俚语中，它指的是官僚体系所特有的那种陈腐、等级森严、自我膨胀的氛围，包括做派的权威和地位游戏。  公司的新办公室。 ## “我不知道她是怎么熬过来的” 在这篇报道的采访之初，我非常紧张。我即将采访一些世界上最聪明的AI人才。我是一个文科生，从未在科技行业工作过，对AI的了解也很有限。 但当我真正开始与来自算法和产品工程团队的年轻专家交谈时，我发现紧张的反而是他们。他们担心如果我听不懂他们的术语会感到尴尬。 于是他们会先把英文翻译成中文，然后再把那个中文翻译成更简单的、我能听懂的中文。 那种想要保护别人的本能令人感动。 在开始采访之前，公司给我的唯一指示是：保护每一个人。 所以我尽量避开过于敏感或可能伤害到人的问题。 即便如此，Ty在一次电话采访中还是没能完全掩饰住一丝情绪的波动。他刚入职时，艰难的适应过程让他非常挣扎。有一段时间，他觉得自己撑不下去了，甚至想过辞职。 后来有一周，在公司会议上，他看到Annie——一个才毕业两年的女生——在无数次挫折和内心动摇之后，终于把一个艰难的项目推进了下去。看到这一幕，他觉得自己也不能放弃。他比她年长，人生阅历更丰富，但在纯粹的耐力和意志力上，他觉得自己更弱。 他说： “我不知道她是怎么熬过来的。” 事实上，想过离开的不止Ty一个人。 Annie也想过。 很长一段时间里，她都在尝试从零到一搭建一条海外业务线，却始终没有真正的突破。更糟的是，其他团队的同事出于好意，直接劝她放弃这项在他们看来毫无意义的努力。 她说，她在Kimi哭过的次数，比为其他任何一家公司、为她曾经交往过的任何一个前男友哭过的都多。 她并不是没有其他选择。她手上已经有一个薪水更高的工作机会。但她说，她就是无法说服自己去为别人工作。她想和张宇桐再谈一次。 那次谈话之后，她决定留下来。 她没有告诉我那次谈话的内容。她只是说：宇桐是我见过的最强的老板，迭代最快，天花板最高。跟着她，我才能提升自己的天花板。 然后安妮重复了那句话： “我不知道她是怎么扛过来的。” 当你收集到足够多的素材时，你会发现某些话在反复出现。而这些被重复最多的词句，往往揭示出一个团队最深层的共同特质。 Bob 曾被杨植麟召回中国，放弃了去美国攻读博士的机会，从公司成立第一天起就加入了进来。如果说有人真正深刻理解这家公司，那一定是他。 当我向他提出我问过所有人的同一个问题——团队最重要的品质是什么——他想了大约两分钟，然后用一个词回答： 韧性。 对一家只有三年历史的公司来说，谈论韧性听起来或许像是一种奢侈。但他是认真的。他说，聪明和勇敢有时是一对反义词。你越聪明，就越清楚地看到风险，也就越容易选择离开。但愚蠢的坚持同样不会成功。所以，只有那些看清了真相、计算出失败的概率、却依然选择继续前行的人，才配得上“韧性”这个词。 公司内部流传着一个故事，叫做“三上反思悬崖”。 2023 年 5 月，Freddie 和他的同事们接到了一项看似不可能完成的任务：让 AI 一次性读取并理解 128K 的上下文，也就是数百页书的篇幅——而当时行业的标准还停留在 4K 左右。 他很快设计了一个名为 MoBA v0.5 的方案，但这个方案需要在主模型训练进行到一半时重写底层训练框架。代价太高，于是这个想法被搁置了。这是第一次上“反思悬崖”。 半年后，他带着版本 1 回来了，这次的设计是从现有模型继续训练。方案在小模型上奏效，但在大模型上测试时遇到了 loss 飙升，反复失败。项目被迫第二次回到悬崖边，又搁置了半年。它甚至错过了公司 20 万字产品的重要节点。但团队并没有被解散。相反，公司发起了一场所谓的“饱和式救援”，从各处召集技术专家共同攻克难题。他们重写了核心逻辑，最终让版本 2 通过了经典的长期上下文“大海捞针”测试。 就在眼看要上线之际，第三次打击降临了。在监督微调阶段，模型在长文本摘要任务上表现不佳，因为训练信号太稀疏。那时巨大的资源已经投入进去。尽管如此，工程师们还是再次回到悬崖边，寻找解决方案，最终通过修改最后几层的注意力机制解决了这个问题。 三次退回。三次归来。 在采访的最后，我问了 Freddie 那个终极问题：你会如何形容这家公司？ 他用两个词回答： 登月。 为什么是登月？ 他引用了约翰·肯尼迪的名言： > 我们选择在这个十年登月，并去做其他事情，不是因为它们容易，而是因为它们艰难。  公司所有的会议室都以音乐组合的名字命名。 ## 天才蜂群 最终，我没有打扰或试图探究几位联合创始人本人。 在外部，他们几乎隐形。他们不喜欢接受采访，对个人名声毫无兴趣。但在内部，他们却无处不在。 在一个极其扁平的组织中，你需要超级大脑坐镇中枢。否则活力会变成混乱。由于中层管理很少，每位联合创始人直接对接大约 40 到 50 名员工，并始终贴近技术和业务的第一线。正是通过这种方式，公司让决策与执行保持一致。 五位联合创始人都来自清华大学。但生理极限依然存在。人的注意力是有限的。管理幅度是有限的。当公司估值达到 1200 亿元人民币、员工超过 300 人时，即使是这些超级大脑，也开始在重压下感到吃力。 而这不仅仅关乎创始人。 这是一场由自我驱动力推动的无限游戏。如果每个成员实际上都在承载 4 亿元的估值，那么每个人都被期望创造非凡的价值。 革命性的变量是工具。 Kimi 实际上并不依靠极端的工作时长运转。员工可以自然醒来。他们不需要每天在办公室熬到天亮。产品团队的 Leo 说，他现在指挥着“一支军队”，指的是 AI 智能体。 想象这样一个场景： Leo 早上 10 点醒来，走进办公室。他的任务是分析过去 24 小时来自五个全球市场的用户反馈，并决定本周的产品优先级。在过去，这需要三个人花两天时间。 现在，他启动了三个智能体。 一个策略智能体扫描 3000 条反馈，筛选出与长上下文中断相关的高优先级需求。一个翻译智能体实时解读日语方言和韩语敬语，并标注真实的情绪强度。一个竞品智能体监控 Cursor 和 ChatGPT 的更新，并生成一份技术对比。 Leo 自己只做三件事。他否决了一条被系统误读为真诚的讽刺评论。他标记了一张包含未发布 UI 的截图。他确认了智能体推荐的前三大需求。 上午11:30，产品需求文档已经完成。与此同时，一个编码智能体已生成了约70%的基础实现，只留下更具创造性的设计工作，留给下午与人类工程师讨论。 人类制定规则。硅基系统执行规则。组织成为一个算法的容器。 在AI原生公司，熟练使用智能体并将其深度嵌入工作流不是可选项，而是工作本身的一部分。 模型不仅是目标，也是工具。 无论是直接提升生产力，还是从根本上改变管理结构，AI的逻辑已经深入这家公司的骨髓。正如公司构建智能体集群，团队本身也开始像一个天才集群：许多独立的天才并行工作，无缝协作。 不过，这种扁平结构有着与生俱来的脆弱性。 当我问起，如果公司从300人增长到3000人，这种模式是否还能持续时，大多数人的回答都很谨慎。历史并不乐观。类似的极端扁平化实验，如合弄制或海尔的内部小微契约单元结构，一旦超过500人左右，往往就会遭遇决策瓶颈。当信息节点过多时，“直接沟通”就开始变成信息过载。 一个更直接的痛点是个人体验到的失重感。 没有了层级来缓冲不确定性，方向感的迷失会被每个人直接感受到。一位最终回归大厂的前员工直言：没有自上而下的OKR和KPI，有些早晨你走进办公室，却不知道自己该做什么。也没有人一定告诉你做得好不好。这种反馈的缺失造成不安全感。它会让人怀念大科技公司明确的汇报线、评审节点和可量化的产出。 那些繁琐的结构，毕竟提供了一样至关重要的东西：确定性的底线。 目标在哪里？什么算完成？绩效如何评判？在大公司里，这一切都清晰可见。 那位人士说，这不是斯德哥尔摩综合征，这是基本的组织物理学。 如果说阿里巴巴像一条精密校准的晋升传送带，字节跳动像一支目标感极强的冷酷战斗军团，腾讯像一所更宽容的职业学院，那么月之暗面就像一片原始森林。 天才或许能找到狩猎之路，普通人可能只是在迷雾中徘徊。 ## 必要的“二向箔” 没有部门。没有头衔。没有考评。 AI原生的组织模式是反官僚的，刻意保持无结构状态。大公司已难以轻易转向这种模式，小公司往往因为过快扩张成传统结构而错失窗口期。这是一场不对称的战争。 在这里，作者引用了《三体》中一个著名的科幻典故。在那个故事里，一个高级文明随意使用了一种叫二向箔的武器，将太阳系从三维跌落至二维。行星、恒星和人类都变成了一幅没有厚度的平面图像。 作者认为，月之暗面正是在刻意将这样一张“二向箔”掷向自己。 不是为了毁灭对手，而是为了将组织扁平化，以追求效率的极致。 没有层级结构的纵向深度，没有部门壁垒的横向隔墙，没有办公室政治的三维纠葛。只有“模型”与“智能”以尽可能简单的形式正面相对。 作者认为，在AI时代，每一家创业公司都被迫向自己掷出这样一张二向箔。一人公司的兴起，反映的是AI原生人才同样在经历代际爆发。如果技术能把组织能力压缩到个体之中，那么管理的中层许多环节就会直接蒸发。组织被压扁。没有留下绕行的纵深。每个人都被迫直面问题本身。 这或许就是商业世界组织演化的一条铁律。 每个人，最终都会被折叠。 一旦人们在同一个平面上暴露，一个人辐射影响五十个人，就不再像是管理奇迹，而成为常态。中心到边缘的距离被重新定义。依赖头衔和OKR作为坐标的人可能瞬间窒息；而天才们，在这张暴露的平面上，可以猛烈地拆解智能本身，“守护者”们则清除噪音与熵增，并不无谦卑地把自己视为拓宽人类文明边界的拓荒者。 然而，从三维跌落二维是不可逆的。 这意味着Kimi无法后退。 每一次战略调整都成为高风险的混沌迭代。竞争对手仍可以在迷宫中缓慢转身，但如果月之暗面试图在规模上鲁莽扩张，可能会在结构上将自己撕裂。这种自我扁平化的行为之所以可以接受，只是因为它服务于某种更激进的东西。 降低组织维度的终点，是提升智能的维度。 只有当模型智能跨越临界阈值，上升到足以摆脱一切碳基组织引力井的高度，月之暗面才能真正碾压竞争对手的组织优势，并为这场不可逆的豪赌正名。 到那时，关于管理幅度或组织架构图的争论都不再重要。那就像追问三体文明栖居于什么维度，而真正的要点在于，它的维度武器早已改写了战争规则。 到那时，“月之暗面”将不再是一个隐喻。 它将成为一个高维光源，照亮智能宇宙的暗面。此前所有的组织之痛，都不过是登月舱穿越大气层时烧蚀脱落的隔热层罢了。 它们要么通过升维成为神一般的存在。 要么在坍缩中被封存。 没有第三条路。 文中使用的所有英文名均为化名。 ## 相关链接 - [Rui Ma](https://x.com/ruima) - [@ruima](https://x.com/ruima) - [58.9万](https://x.com/ruima/status/2039245985520681257/analytics) - [升级至高级版](https://x.com/i/premium_sign_up) - [2026年4月1日 下午3:38](https://x.com/ruima/status/2039245985520681257) - [58.94万次浏览](https://x.com/ruima/status/2039245985520681257/analytics) - [查看引用](https://x.com/ruima/status/2039245985520681257/quotes) --- *导出时间: 2026/7/22 09:40:16*
1 100 小时深入基米：揭秘 Moonshot AI 的极客文化与生存法则 本文通过100小时的内部观察，深入剖析了Moonshot AI（Kimi）独特的企业文化与生存状态。文章记录了公司如何在DeepSeek崛起的压力下调整战略，强调了“模型能力”至上的原则。作者揭示了该公司偏爱具有“概括能力”和“品味”的天才型员工，实行无KPI、无层级的扁平化管理，并详细描述了年轻员工在高强度跨领域工作中的成长与焦虑。这不仅是一家初创公司的生存侧写，也是对AI时代人才组织模式的深刻探讨。 技术 › LLM Rui Ma 2026-04-01 Moonshot AIKimi公司文化DeepSeek人工智能职场AGI访谈创业模型训练
8 8款前沿大模型能力横评：10个维度50项测试全解读 本文对8款前沿大模型进行了为期两天的深度横评，通过10个维度、50个细项的两轮测试，对比了模型在一次性复杂任务与拆解任务稳定性上的表现。文章详细解读了测试方法、评分标准及各模型优劣势，最终得出了适用于真实工作流的参考结论。 技术 › LLM huangserva 2026-08-17 大模型评测LLM横评GPTClaudeKimiDeepSeekQwen性能测试工作流
2 2026年普通人都能上车的AI风口：DeepSeek 创始人梁文锋3小时演讲精华提炼 文章提炼了DeepSeek创始人梁文锋关于AI未来的演讲精华，强调“克制”的重要性。梁文锋指出，想拿得多的人会被想拿得少的人打败。他为AI发展制定了清晰的路径，并建议普通人不要盲目追逐风口，而应注重提升“把话说清楚”和“长期沉淀”的能力，通过行业深耕和AI应用找到属于自己的机会。 技术 › LLM Gloria 2026-07-24 DeepSeek梁文锋人工智能职业发展创业思维链智能体Coding AgentAI应用克制
与 与杨植麟的第二次对话：站在无限的开端 本文是张小珺对月之暗面创始人杨植麟的第二次深度访谈，发生在Kimi K2模型发布后。文章深入探讨了AI如何通过编码能力突破“缸中之脑”，走向Agent与自我进化，以及杨植麟对技术挑战与创业心境的思考。 技术 › LLM 张小珺 2026-07-22 杨植麟Kimi访谈Agent技术哲学DeepSeek自我进化创业MoEThe Beginning of Infinity
A A Conversation with Yang Zhilin of Kimi: Advancing Toward the Endless, Unknown Snow Mountains 本文是对月之暗面创始人杨植麟的专访，回顾了他从硅谷回国创业、专注AGI的历程。文章探讨了基础模型公司面临的资本、人才竞争压力，以及Kimi在商业化与理想主义之间的平衡。 技术 › LLM 张小珺 Xiaojun Zhang 2026-07-19 杨植麟Kimi月之暗面AGI访谈创业基础模型商业化资本技术理想主义
D DeepSeek新论文解读：用“手指着图片思考”的多模态推理框架 DeepSeek发布新论文，提出一种名为“视觉原语”的多模态推理框架。不同于业界卷分辨率的趋势，该模型通过在思维链中嵌入边界框和坐标点，模拟人类用手指着图片思考的过程，有效解决了语言在空间指代上的局限。DeepSeek-ViT配合LLM将视觉Token压缩至竞品的十分之一，但在空间推理、计数和迷宫导航等任务上性能超越GPT-5.4等顶尖模型。 技术 › LLM 向 向阳乔木 2026-05-01 DeepSeek多模态视觉推理论文解读Transformer人工智能计算机视觉LLM
L LMArena 最新排名：文心 5.1 预览版国产登顶，文本能力仍是核心 本文基于 LMArena 最新文本排行榜更新，探讨了文心 5.1 Preview 凭借 1476 分拿下国内第一、全球第 13 的成绩。作者分析了为何在 DeepSeek V4 和 GPT-5.5 主导的时代，文本能力依然是大模型代码生成与逻辑推理的“基本盘”。文章指出文心通过「多维弹性预训练」技术实现了高性价比与高性能，验证了夯实文本底座的重要性。 技术 › LLM huangserva 2026-05-01 LLM文心一言DeepSeekLMArena模型排名文本生成百度人工智能
人 人类最后的职位：上下文耕作 文章探讨了 AI 时代企业与人类角色的根本性转变。随着“代理微公司（AMC）”的兴起，传统的层级制企业将被拥有“公司大脑”的小型团队取代。人类的核心工作不再是执行或决策，而是作为“上下文耕作者”，为 AI 提供高质量的信息环境。文章指出，构建可组合的全球性上下文基础设施将是未来的万亿美元级机会。 技术 › Agent brett goldstein 2026-04-25 AgentAMC公司大脑上下文未来工作创业组织架构LLM人工智能行业趋势
推 推理是训练的 12 倍，可一块卡生成时只跑出 5% 的算力 文章基于 YC Paper Club 播客，剖析 LLM 推理成本真相：按两年模型寿命摊算，推理总算力约为训练的 12 倍。decode 阶段 MFU 仅 5%，瓶颈在显存带宽而非算力。DeepSeek 通过 KV cache 保留 72 小时、radix attention、零开销调度等手段将缓存命中率做到 96%，实现 38 美元一卡时、80% 毛利。文章还讨论了 SSM 等被低估的模型结构、服务商质量参差（同一模型不同服务商差距超开源闭源差距）以及开源推理生态的规模，指出推理成本优化才是 LLM 竞争的胜负手。 技术 › LLM Saito 2026-10-26 LLM推理优化DeepSeekSGLangKV缓存MoE量化开源模型推理解码算力成本
K Kimi K3.1 Pro 第一手测评：与 GPT-6 Astra 硬刚游戏开发 月之暗面发布 Kimi K3.1 Pro，3万亿参数MoE架构、100万token上下文、开源权重，API输出价格仅为GPT-6 Astra的五分之一。作者从跑分和实测两方面评估：官方xdev-bench-hard编程成功率64.5%排第二，幻觉测试同样第二，第三方智能指数全场第四、开源第一。实测中，作者用相同提示词让K3.1 Pro与GPT-6 Astra分别开发3D跳一跳和植物大战僵尸游戏，两者均一次成型无bug。Kimi耗时约10分钟但会自我验证和修正，Astra快近5倍且细节更丰富。结论：K3.1 Pro编程能力接住闭源旗舰，价格是杀手锏，适合重度Agent用户。 技术 › LLM 程序员小灰 2026-10-25 KimiK3.1 Pro模型测评开源模型编程能力GPT-6 Astra游戏开发月之暗面MoEAgent
K Kimi K3.1 Pro 来了：离最强只差 0.7 分，API 反而降价三分之一，权重 11 月 4 日白送 月之暗面发布 Kimi K3.1 Pro，基于 K3 做 post-training 升级，2.8T 总参数、104B 激活（MoE）、1M 上下文、多模态。Terminal-Bench 4.0 得分 65.7，超越 GPT-6 Astra（58.2），仅落后 Claude Opus 5.5（66.4）0.7 分，是开源模型离闭源最强最接近的一次。API 价格输入 $2/输出 $10，比上一代降价三分之一，权重将于 11 月 4 日开源。文章拆解了 14 项评测表现，指出其在编码、Agent、金融分析领域表现出色，但科学计算类任务偏科严重（38.1 vs 68.1），并给出小白使用入口与避坑清单。 技术 › LLM 丰收 2026-10-24 Kimi月之暗面大模型开源模型MoEbenchmarkAPI定价Agent编程能力模型评测
K Kimi K3.1 Pro来了，国产模型审美的新巅峰 月之暗面发布旗舰模型Kimi K3.1 Pro，参数规模2.8T（MoE，104B激活），支持100万Token上下文，继续开源。API价格较K3砍半，仅为GPT-6 Astra的五分之一。跑分仅次于Claude Opus 5.5和GPT-6 Astra，Terminal-Bench 4.0达65.7%，OfficeQA Pro v2获SOTA，幻觉率从23.4%降至8.3%。作者重点测试了视觉方向：3D建模能力已可与GPT-6 Astra正面竞争，包括网页3D果冻、高精度机械键盘、根据户型图与装修图生成Three.js第一视角漫游网站、结合平面图与手持视频还原办公室3D模型。视频动效方面提升巨大，可一轮直出动效作品集、宣传片及科普视频，虽距Opus 5.5仍有差距，但已属国产模型巅峰水平。 技术 › LLM 数字生命卡兹克 2026-10-24 KimiK3.1 Pro月之暗面大模型开源3D建模视频动效模型评测国产AIMoE
