Post-90s Tsinghua Graduate's AI Startup Reaches $2.5 Billion Valuation in One Year
-
Currently developing a general multimodal model in secrecy, with plans to release it within this year.
At a time when OpenAI's text-to-video model Sora has captured global attention, this domestic AI company founded by post-90s Tsinghua graduates continues to attract investments from prominent institutions.
Today, Moonshot AI reportedly completed a new funding round exceeding $1 billion, with investors including Sequoia China, Xiaohongshu, Meituan, and Alibaba, along with follow-on investments from existing shareholders. The post-money valuation reached approximately $2.5 billion.
Just before the Spring Festival on February 3, Moonshot AI was reported to be raising $200 million in funding, jointly invested by Ant Group and Alibaba Group, with a pre-money valuation estimated at $1.5 billion. Regarding the "latest round of funding," Moon's Dark Side told Lieyun Network, "Thank you for your attention! The company is not in a position to comment on specific financing details at this time. Moon's Dark Side has always been committed to promoting the underlying technological advancements and product innovations in China's AGI field, and we will continue to align our capital strategy with the company's development stage. We look forward to sharing more good news with everyone in the future."
In fact, Moon's Dark Side, founded in April 2023, has been "rarely explicit" about its funding disclosures. Regarding the first round of financing, which was reported just two months after its establishment, founder Yang Zhilin "corrected" in October 2023 that the company had received nearly 2 billion yuan in investment from renowned institutions such as Sequoia Capital, Today Capital, and LiSi Capital.
This remains the only accurate funding disclosure since Moon's Dark Side was founded nearly a year ago.
Tsinghua Elite Entrepreneurship, Assembling an All-Star Team with Fellow Alumni The reason why Dark Side of the Moon received investments from top VCs at its inception is closely related to the "top student" identity and extensive experience of its post-90s founder, Yang Zhilin.
During his studies at Tsinghua University, Yang Zhilin studied under Professor Tang Jie, the leader of the Knowledge Engineering Group (KEG) at Tsinghua's Computer Science Department, academic vice president of the Beijing Academy of Artificial Intelligence, and head of the WuDao project. He graduated with full marks in all programming courses and ranked first in his class.
In 2015, Yang Zhilin entered Carnegie Mellon University's Language Technologies Institute (LTI), where he pursued his Ph.D. under the guidance of Ruslan Salakhutdinov, Apple's AI head, and William W. Cohen, Google's Chief AI Scientist.
After graduation, Yang Zhilin worked at Google Brain and Meta (Facebook) AI Research, where he became the first author of Transformer-XL and XLNet. The XLNet model outperformed Google's BERT in 18 natural language processing tasks, making it one of the hottest international cutting-edge models in the NLP field at the time. According to incomplete statistics, Yang Zhilin has published more than 20 papers in top-tier computer conferences such as ICLR, NeurIPS, ICML, ACL, and EMNLP, with his research accumulating over 17,000 citations on Google Scholar.
Currently, Yang Zhilin is also an assistant professor at the Institute for Interdisciplinary Information Sciences at Tsinghua University. His research focuses on large-scale pretraining, natural language processing, natural language understanding and generation, few-shot learning, zero-shot learning, and multimodal learning.
As a post-90s scholar, Yang Zhilin is highly renowned in the field of large models: his name and presence can be found everywhere, from Recurrent AI, Zhipu AI, to the Beijing Academy of Artificial Intelligence.
Additionally, Yang Zhilin and his team have served as core R&D members in the development of large models such as Google Bard, Gemini, Einstein, Pangu, and Wudao. They have invented milestone achievements in the AI field, including TransformerXL, XLNet, RoPE, Detectron2, and Group Normalization, which have been adopted by models like Google PALM and LLaMa. According to Tianyancha App, Yang Zhilin holds 78.97% of the shares in Moon's Dark Side, giving him absolute control. The entrepreneurial partners around him are equally formidable and not to be underestimated.
Zhou Xinyu, co-founder of Moon's Dark Side, holds 10% of the company's shares. He, along with Yang Zhilin and Zhang Yutao, were classmates in the 2011 undergraduate class of the Department of Computer Science and Technology at Tsinghua University. In his senior year, Zhou Xinyu joined Megvii for an internship, which met all his criteria, and officially joined after graduation. His work focused on algorithm mass production, aiming to significantly improve the efficiency of algorithm production.
As the third-largest shareholder with a 5.96% stake, Wu Yuxin, another co-founder of Moon's Dark Side, graduated from Tsinghua University and Carnegie Mellon University. He was nominated for the Best Paper Award at the 2018 European Conference on Computer Vision (ECCV). At the GeekPwn International Security Geek Competition in October 2018, IYSWIM was the only team among six participants to successfully crack a facial recognition algorithm. Wu Yuxin participated as part of the IYSWIM team and explained that he used Google's FaceNet open-source code model to breach the algorithm (he registered under his own name, with a teammate not present at the event). Additionally, Yang Zhilin's fellow student Zhang Yutao currently holds a 5% stake in the company. Public records show that Zhang Yutao completed both his bachelor's and master's degrees in the Computer Science Department at Tsinghua University. His research focuses on heterogeneous data fusion and knowledge graph construction, with multiple papers published in top-tier computer conferences such as KDD and CIKM. As the technical lead, he participated in the development of the AMiner platform for scientific big data analysis.
After Large Language Models, Secretly Developing General Multimodal Models
With its elite team and profound expertise, MoonShot AI announced a breakthrough in the 'long text' domain within six months of its establishment in October 2023.
According to Yang Zhilin, to address the 'application challenges caused by the limited input length of large models,' MoonShot AI officially launched its first large model, moonshot, which supports input of up to 200,000 Chinese characters, along with the intelligent assistant product Kimi Chat powered by this model. Subsequently, he provided detailed examples of Kimi Chat's practical applications. Taking the entire book "The Moon and Sixpence" as an example, Kimi Chat can read along with users, helping them better understand and apply the knowledge from the book:
Compared to current large model services primarily trained on English, Kimi Chat possesses strong multilingual capabilities. For instance, it shows significant advantages in Chinese, practically supporting contexts of about 200,000 Chinese characters - 2.5 times that of Anthropic's Claude-100k (approximately 80,000 characters in actual tests) and 8 times that of OpenAI's GPT-4-32k (approximately 25,000 characters in actual tests).
Meanwhile, through innovative network architecture and engineering optimizations, Kimi Chat has achieved lossless long-range attention mechanisms under hundreds of billions of parameters, without relying on performance-damaging "shortcut" solutions like sliding windows, downsampling, or smaller models. On January 26th this year, Kimi Chat launched its latest 'v1.3 Spring Edition': the base model capabilities have been comprehensively upgraded, featuring online search capabilities, contextual learning, literary creation, language translation... The mini-program version of Kimi Assistant now supports Chinese and English voice input.
Meanwhile, OpenAI's text-to-video model Sora has been continuously drawing attention due to its breakthrough in generating videos longer than one minute, along with the highly realistic and high-quality demonstration videos.
According to multiple reports, Moon's Dark Side is secretly developing a general multimodal model, expected to be released within this year.
In other words, after achieving phased progress in the text-based large language model field, Moon's Dark Side will soon compete with OpenAI and other domestic and international counterparts in the multimodal model domain, which primarily focuses on images and videos. Moon's Dark Side's frequent moves in the capital market appear to be preparations for higher training costs and greater demands for funding and talent.
Taking team size as an example, Moon's Dark Side had around 50 employees in October 2023, but the company's workforce has now exceeded 80.
In the future, we will continue to monitor whether Moon's Dark Side can secure a place in the field of "multimodal models."