Skip to content

X5H Info Portal

Bilingual coverage of industry news, policy and practice

繁中EN
Business

Original Yang Zhilin broke his wrist: Kimi’s darkest moment and expedition

2026-10-06 X5H 編輯部 29 min read

Text | Yang Wanli

Yang Zhilin, who dreamed of becoming a rock star when he was a boy, would never have imagined that the AI ​​company he founded many years later would be named after a classic rock album.

Only one side of the moon always faces the earth, and that hidden “dark side” symbolizes the unknown boundary that has not yet been glimpsed. In April 2023, Yang Zhilin named the startup company “The Dark Side of the Moon”, which was taken from the classic album “The Dark Side of the Moon” by the rock band Pink Floyd. He hid his pursuit of general artificial intelligence in this romantic name.

At that time, perhaps not many people took this “rock meme” seriously, until Kimi of The Dark Side of the Moon successfully broke out of the circle, and Yang Zhilin and his startup company came into the spotlight.

In early September this year, media reports stated that Dark Side of the Moon was considering dual listing in Shanghai and Hong Kong. From international capital to the domestic market, if this plan is successfully implemented, it will become the first company in the domestic large-scale model track to simultaneously promote the listing of Hong Kong stocks and the Science and Technology Innovation Board.

The road behind the halo has not been smooth: Facing the impact of competing products, the company once fell into the darkest moment of development. It was Yang Zhilin who made an almost decisive decision to lead the team out of the trough and embark on the expedition to general AI.

From an Olympiad boy, a Tsinghua student, a CMU doctor, to the helm of a large model startup, what kind of path has Yang Zhilin taken? What setbacks and breakthroughs has Dark Side of the Moon experienced in the fierce AI race, and what long-term ideals does it have?

From competition boy to big model light chaser

Yang Zhilin could have enjoyed life with a high salary overseas, but he chose to start a local large-scale business from scratch.

Yang Zhilin studied at Shantou Jinshan Middle School. In high school, Yang Zhilin, who had no programming foundation, was selected into the Informatics Olympiad competition training class. Later, by winning the first prize in the Guangdong Division in the National Youth Informatics Olympiad League, he was recommended to Tsinghua University.

After Yang Zhilin entered Tsinghua University, he initially majored in thermal engineering and switched to computer science in his sophomore year.

When Yang Zhilin was a boy, he dreamed of becoming a rock star or a wandering poet, and he also formed a campus rock band. After being influenced by a novel, he chose to be a coder. Upon graduation, he passed all programming courses with perfect scores. Yang Zhilin’s undergraduate life was relatively smooth. His biggest regret was that his band failed to win the championship in the original competition.

In 2015, Yang Zhilin went to the United States to study for a doctorate. In 2019, he obtained a doctorate in computer science from Carnegie Mellon University. In 2019, Yang Zhilin, as the first author, jointly published the deep learning and natural language processing model XLNet with Google and Carnegie Mellon University, emerging in the field of deep learning.

At that time, he had admission tickets to top overseas research institutes and major technology companies, and his future was bright. But in Yang Zhilin’s view, those who stay in large Silicon Valley factories and top research institutes, with generous remuneration and decent work, are ultimately just senior technical employees, and what he wants is to completely dominate the research and development of large models and commercialized products. Yang Zhilin once said in an interview, “If I don’t start my own business, I will regret it for the rest of my life.”

In the spring of 2023, Yang Zhilin founded Moonshot AI. With the yearning for general artificial intelligence, Yang Zhilin and his team embarked on an entrepreneurial journey. According to the official website of The Dark Side of the Moon, they chose this name because they believe that truly significant things are often at the end of human sight.

The first principle of Dark Side of the Moon is to convert energy into intelligence. The transformation they seek is not a cold computing competition, but an eternal question about the upper limit of intelligence – how deeply a model can think, how far humans can go.

Decisively “cut off your wrist” at the darkest moment and launch an “expedition” into the future

The rise of Dark Side of the Moon was not a smooth road. The brutal competition in the market soon gave this start-up company a severe test.

In October 2023, Kimi was officially launched. It became an instant hit with its ability to understand 200,000-word ultra-long texts. It became the first business card of Dark Side of the Moon to the market, and also kicked off the domestic competition for large model long texts. After multiple rounds of version iterations in 2024, Kimi user traffic has experienced explosive growth.

According to media reports, within ten months, Kimi’s monthly active users soared from 3 million to 21.01 million, an increase of 700%. Dark Side of the Moon successfully received US$1 billion in financing led by Alibaba, and its valuation exceeded US$2.5 billion in one fell swoop. It topped the list of domestic large model unicorn valuations in 2024 and became an “AI upstart” sought after by capital. However, the industry is changing. In early 2025, with its low training cost and high performance, DeepSeek was rated as a “national sports-level breakthrough” by the outside world. This wave not only shocked American technology manufacturers, but also impacted Kimi, which targets C-end users.

According to media reports, in January 2025, the number of DeepSeek users increased to 125 million, of which 100 million visiting users were achieved within seven days. According to QuestMobile data, DeepSeek’s monthly active users reached 190 million in March 2025, while Kimi’s monthly active users were less than 20 million during the same period, a difference of more than 9 times.

Under the squeeze of strong opponents, the track pattern of the AI ​​”Six Little Dragons” has changed drastically: many peers either choose to rely on head strength or retreat to subdivided tracks to seek differentiated survival. The spring of 2025 became the darkest hour of the Dark Side of the Moon.

At this critical stage of development, Yang Zhilin made a very bold decision to stop marketing, take back resources, and focus on improving the underlying performance of the model while exploring product functional innovation and finding a path to differentiated competition.

According to media reports, a Kimi internal researcher once revealed that at the reflection meeting at the beginning of 2025, the company team put forward some quite radical suggestions. Unexpectedly, Yang Zhilin’s subsequent actions were even more radical than expected, such as no longer updating the K1 series models and concentrating resources on basic algorithms and K2.

The hibernation of shrinking the front line will soon pay off. In June 2025, Kimi’s in-depth research function will start internal testing. Models are no longer just conversational tools, they start to think like humans, ushering in a new era of “models as agents”.

In July 2025, the trillion-parameter model K2 was officially released and open sourced, which was the most important turning point for Dark Side of the Moon; two months later, Kimi launched the Agent mode, codenamed “OK Computer”.

By the end of 2025, Yang Zhilin announced in an internal letter a $500 million Series C financing. At that time, Dark Side of the Moon’s technology, products and business had achieved comprehensive breakthroughs. In the year 2025, Dark Side of the Moon has completed the “low start and high move” rebound from the trough, and is ready to embark on an expedition towards general AI. Yang Zhilin set a goal in an internal letter at the end of 2025: in 2026, we will create a unique and undefined large language model.

In 2026, Dark Side of the Moon will continue to iterate model products at a high speed. In January, Dark Side of the Moon released Kimi K2.5, a model that ranked high on multiple lists. On August 4, 2026, Dark Side of the Moon Kimi issued a notice to its API developers, announcing that the kimi-k2.5 model and moonshot-v1 series models will be offline on August 31, 2026.

Just half a month ago, Kimi K3 was released on July 16, 2026. With 2.8 trillion parameters and 1 million token context windows, it is the world’s largest open source model.

After the release of Kimi K3, it attracted widespread attention in the global technology field. The model ranked among the top three in the world and led other mainstream large models, and narrowed the gap with the leading models in the United States. This alarmed Musk and Wall Street, and was no less than a “DeepSeek” moment.

Technological breakthroughs drive accelerated commercialization. As of mid-June 2026, Dark Side of the Moon ARR (annual recurring revenue) exceeded $300 million.

Actions at the capital market level are also intensive: starting from the end of June 2026, the company has completed pre-IPO work such as shareholding reform and board of directors filing. There are constant rumors of dual listing in Shanghai and Hong Kong, and it is getting closer and closer to landing on the secondary market.

The Dark Side of the Moon wrote this on its official website, “General artificial intelligence does not belong to any one company or any land. It will eventually flow to every corner of the world. We are willing to be the first group of people to light it up with the persistence of open models and technical idealism, as well as the sincerity to create together with millions of users.”

Related reading

About Us · 關於我們