Tsinghua talent + rock drummer—this Chaoshan guy is making Silicon Valley restless

Author: seed.eth; Source: Bifan

Late at night on July 16, 2026, when Yang Zhilin’s Kimi K3 model went live, no one expected what would happen next.

On the first day, it topped the Arena AI front-end code competition arena with 1,679 points, beating Claude and GPT. Overseas, some people dubbed it a “DeepSeek 2.0 moment.”

On the second day, Musk wrote “Impressive” on a social platform, and immediately announced that his own new model with 20 trillion parameters might surpass Kimi. The Dark Side responded: “Welcome Musk to the ‘20 trillion+ club’.”

On the third day, demand for Kimi exceeded expectations, pushing the cluster close to its capacity limit. The team issued a late-night announcement to pause new user subscriptions.

On the fourth day, officials from the White House publicly accused The Dark Side of “stealing technology” and “circumventing chip export controls.”

On the fifth day, Nasdaq opened down 1.8% due to factors including the K3 release, while the Philadelphia Semiconductor Index fell 5.2%.

On the sixth day, Bloomberg reported: K3 is believed to possibly narrow the AI gap between China and the US. Some scholars pointed out that the gap might shrink to two to three months.

On the seventh day, Hu Xijin posted a reminder to Yang Zhilin: “For now, don’t go to the United States.”

In seven days, a post-90s person from Shantou made Silicon Valley, the White House, and Wall Street all stumble.

Code is the tool, rock is the bone

Yang Zhilin, born in 1992 into a typical family in Shantou, Guangdong, had a path that was different from the start.

As a teenager, he had two hobbies: rock music and coding. He attended Shantou Jinshan High School. With no programming background at all, he was selected into an informatics olympiad training program. Most of his classmates started writing code as early as junior high. He started late—so late it was almost absurd. But just a year later, he won first place in the Guangdong division of the National Youth Informatics League, earning a Tsinghua University recommendation.

But he insisted on proving he wasn’t just a “recommended student.” He went through the self-enrollment selection again and barely passed, then decided to take the national college entrance exam as a regular high school student—scoring 667 points to become the top science student in Shantou. Three times admitted to Tsinghua, making him a local legend.

In 2011, Yang Zhilin (center), top science student in Shantou’s college entrance exam. Air China provided him with free flight tickets. (Photo source: Hualong Chaoshan)

After entering Tsinghua, he was reassigned to Thermal Energy Engineering—colloquially “stoking the boiler.” In his sophomore year, he made a decision that baffled others: he transferred majors to the Computer Science department. The reason was very “rock”: he liked it. And that liking came from a novel by Haruki Murakami—a role in the novel, a programmer who writes code late at night to turn technology into reality, filled him with longing.

Transferring majors meant he had to make up all of everyone’s freshman programming courses. But in the end, he graduated with the top score in his class, with 90% of his professional course grades above 95. At the same time, he formed a rock band at Tsinghua called Splay, serving as the drummer and lyricist. The band’s name came from the data structure “Splay Tree”—a clever pun. In the 2014 Tsinghua University school song competition, they won the “Best Original Song Award.”

(Top of credits GPA, countless papers)

Years later, when Yang Zhilin looked back on his biggest regret in his undergraduate years, he said: “My band couldn’t win the championship in an original music competition—we only got a Best Original Song Award.”

Entrepreneurship, written into a Chaoshan person’s genes

In 2015, Yang Zhilin graduated with a first-place performance in Tsinghua’s Computer Science program and went to Carnegie Mellon University for a PhD, studying under Ruslan Salakhutdinov, Apple’s first AI director. He finished four years later—two full years faster than the usual six.

During his PhD, he did two works that later became frequently cited: Transformer-XL and XLNet, with a total of cited more than twenty thousand times. Transformer is the underlying backbone of all large models today; his work was akin to making key node improvements on that backbone. Because of this, he became the most-cited researcher in China’s NLP field under age 35.

After K3 was released, Yang Zhilin became famous, and many asked: why didn’t he stay in the United States? Some speculated it was due to a visa; others said he didn’t get selected in H-1B.

The rumors spread too widely that his advisor, Russ Salakhutdinov, had to step out to clarify: the fact is, with Yang Zhilin’s level, staying in the US would have offered countless opportunities close to graduation. Apple wanted to hire him; Google and Meta also offered opportunities. Stanford and MIT asked whether he would like to do a postdoc. Even an Apple executive gave him a position in an office in Beijing. Salakhutdinov said: “I remember he told me that if he didn’t even have the chance to try entrepreneurship, he would regret it for his whole life. I respect his decision, and he’s right.”

In February 2023, Yang Zhilin began focusing on the first round of fundraising.

Later, he recalled that it was an extremely narrow window: “If we delayed until April, basically there would be no chance. But if we did it in December 2022 or January 2023, there would also be no chance—back then there was the pandemic, and everyone didn’t react, so the real window was one month.” While he was still in the US, one night he did precise calculations. After finishing, he felt that at least $1oo million would be needed within a few months. At that time, not many people were raising funds, and many believed he couldn’t raise that much. Later it proved possible—indeed, even more.

On April 17, 2023, The Dark Side was established in Beijing. The company name comes from the album “The Dark Side of the Moon” by Pink Floyd, which he had listened to for many years. Among the co-founders was Zhou Xinyu, the lead guitarist of Tsinghua’s Splay band that year. From a band to a company, the two went from playing music to building models.

Six months later, the Kimi AI assistant officially launched. It focused on a direction that no one dared to do at the time—ultra-long context. A lossless input of 200 thousand Chinese characters represented the longest context length supported among global large-model productized services at that time. As a comparison, Claude then supported about 80 thousand characters, while GPT-4 had only about 25 thousand characters. Yang Zhilin’s judgment was that context length was key to unlocking new application scenarios—from analyzing entire codebases to long-term assistants that wouldn’t forget critical information. While others competed on general leaderboards, he bet that long texts would be the next entry point.

He bet right.

In November 2023, Kimi officially opened its service to the general public. In February 2024, The Dark Side completed more than $1 billion in its A+ round of financing. Alibaba led the round, with follow-on investment from Sequoia China, Xiaohongshu, and Meituan. Post-investment valuation was about $2.5 billion. This was the largest single-round funding amount for domestic AI large-model companies at the time. Less than a year after its founding, The Dark Side became one of China’s most highly valued AI unicorns on the large-model track. In the same year, in March, Kimi increased its context length from 2 million characters to 20k characters. Month-over-month, user visits grew by 321.58%, and it became so hot that it even caused downtime.

From founding in April 2023 to becoming an AI unicorn, Yang Zhilin took less than a year.

Yang is often compared with Liang Wenfeng, founder of DeepSeek, and the two are frequently put side by side. Two fellow Guangdong natives, and they built two of China’s most sought-after AI companies.

But their underlying temperament differs. Liang Wenfeng took a local path—he entered Zhejiang University at age 17, started from quantitative investing, and put AI money he earned himself into it. Yang Zhilin took the academic route.

On business models, Kimi is a financing-driven star company, close to product; DeepSeek is more like a research institution, running long-term on its own funds. In industry discussions, the model capability is what people talk about most.

The narrative difference is even clearer, and some summarized it very directly: Liang Wenfeng captured the track with three keywords—open source, low price, and localization. Yang Zhilin brought globally leading model capabilities, but fought on a battlefield defined by others.

However, by 2026, the two routes started to converge. Kimi increasingly emphasized underlying capabilities, while DeepSeek began to pay attention to reasoning efficiency and productization—the boundaries became increasingly blurred.

China’s AI talent turns Silicon Valley’s heads

Not only Yang Zhilin and Liang Wenfeng stood under the spotlight.

In the same month, Zhipu’s market value at one point surpassed HK$20k, exceeding Xiaomi. Founded by Táng Jié, a professor from Tsinghua, the company had just listed on the Hong Kong Stock Exchange half a year earlier. At age 49 this year, Tang Jie is the oldest among the four founders. In July, he sent an internal letter announcing that over the next two years, the company would concentrate on AGI foundational research and not pursue short-term application monetization. Interestingly, Yang Zhilin was Tang Jie’s student when studying at Tsinghua. Now, teacher and student stand on the same stage, becoming each other’s most familiar “opponents.”

Yan Junjie of MiniMax, focusing on mixture-of-experts models and multimodal technology. Almost at the same time, he also sent a company-wide letter, stating that he would not take a salary until AGI arrives, and he would put forward 4% of his personal shares to incentivize the team. At that time, MiniMax’s stock price had already fallen from HK$1,330 to HK$297.

The four people—Liang Wenfeng at 41, Yang Zhilin at 34, Tang Jie at 49, and Yan Junjie at 37—are called by the market the “Four Heavenly Kings of domestic AI.” Combined, the valuations of the AI companies they control exceed 1.1 trillion yuan. Their backgrounds are different—quantitative trading, computer science, AI research—all top academic prodigies. Their technical routes also each go their own way: DeepSeek cuts costs, Kimi challenges the US frontier models, Zhipu bets on AGI foundational research, and MiniMax expands multimodal applications.

But what puts them all on the same stage is one shared fact: China’s large models are at the front of the global AI competition.

In the past few years, China’s AI narrative was mostly about “catch-up.” In early 2025, DeepSeek shocked the world with a low-cost model—that was the first “DeepSeek moment.” And when Kimi K3 arrived, the industry called it a “DeepSeek 2.0 moment”—a high-performance Chinese large model appearing in the global AI era narrative as a competitor. OpenAI CEO Greg Brockman publicly admitted that estimates he saw showed China might be behind the US in model development by only about four months.

What also makes Silicon Valley tense is the open-source approach commonly adopted by Chinese companies. On OpenRouter, the six most popular models all come from Chinese companies. Global developers can download them for free, deploy independently, and modify flexibly, greatly lowering the application barrier for AI technology. Le Monde pointed out that mainstream Chinese large models generally use an open-source approach, sharply contrasting with the closed-source paid models of US top companies. Some analysts believe that China’s AI industry successfully turned the “chip weakness” into an “algorithm efficiency advantage”—even though the number of GPUs obtained was far less than that of US competitors, they instead practiced extreme engineering optimization.

And propping all of this up is an increasingly young face.

At the 2026 World AI Conference, many people noticed an obvious change: founders from the post-00s and post-90s cohorts accounted for a very high proportion.

Guo Yang, co-founder and CTO of Mianbi Intelligent, was awarded “Annual AI Figure” at CCTV’s “2026 China AI Gala.” Tianwu Technology’s protein design agent “Xiaowu” became one of the event’s “treasured exhibits.” The developer was “post-00s” PhD student Tan Yang. In early 2026, Chen Boyuan—still not graduated—founded Inverse Matrix. In its first round, it received over $1B in investment from Grit Capital and a Peking University funding group, and its valuation had already exceeded 5 billion yuan.

From Yang Zhilin to Chen Boyuan, China’s AI narrative is being rewritten by an even younger generation. Different paths, different directions, but with a shared characteristic—they care very little about established rules, and they care even less about what others think.

(Photo source: People Magazine)

The employees of The Dark Side say the company has no department walls, and in a sense even has no departments. Yang Zhilin’s email signature is only four characters: “Direct communication.” If something needs others’ cooperation to be completed, just contact the person directly—no approvals, no coordination, and no meetings. More than ten employees have said that they’re more used to working with AI, because AI is more reliable and simpler. In group chats, everyone is lively; yet when they meet face-to-face, everyone is quiet. An employee used a word to describe this atmosphere: shy.

Most brands want a story, but Kimi employees will remind visitors: don’t write Pink Floyd, and don’t mention the piano sitting at the entrance to the office.

Going back to the “Longhu ranking” at Shantou Jinshan High School in 2011, the line Yang Zhilin wrote for junior fellows and classmates was: “Believe that things will work out naturally. You can do anything you set your mind to.”

Looking back after more than a decade, that sentence is not only his own footnote—it also feels like a prophecy for the whole generation. They didn’t take shortcuts, and they didn’t detour. They did the things they should do; naturally, the world heard it.