Haidian has spawned another dark horse in the AI field: the 27-year-old Tsinghua professor has rolled out the first RSI model.

The shareholders behind it have gathered top financial investors, industrial capital, Tsinghua-affiliated funds, and the most influential state-owned capitals in Beijing, Shenzhen and Shanghai.

Last month, ChinaVenture mentioned a new trend in the current AI circle in its article "Zhipu Steps In, Investors Scramble to Invest in RSI | Frontline" — RSI (full name: Recursive Self-Improvement). In simple terms, it means allowing AI to generate data, design tasks, formulate rules, conduct evaluations by itself, and adjust and iterate according to the results to complete its own evolution.

This is one of the hottest cutting-edge directions for the next-generation large models. Foreign players such as OpenAI and Anthropic, and domestic players such as Zhipu and Unisound have successively announced their layout in this direction. The article also mentioned that a company named Chaoyan Intelligence is currently a top favorite among domestic RSI entrepreneurial teams.

I first came into contact with Chaoyan Intelligence back in September, when the company had just completed its angel round and angel+ round of financing with a total cumulative amount of nearly 400 million yuan. Among them, the angel round was jointly led by IDG Capital, StarLink Capital and CrystaTech, with participation from Daxun Capital, Infinity Fund, Initial Capital and Yunxiu Capital; the angel+ round was jointly led by Zhongguancun Science City Fund, Shenzhen Venture Capital and Shanghai Future Industry Fund.

Just three months after its establishment, its shareholders have already gathered top financial investors, industrial capital, Tsinghua-affiliated funds, and the most influential state-owned capitals in Beijing, Shenzhen and Shanghai, which fully demonstrates its strong fundraising appeal.

Now, only one month after the financing was completed, the company has released new news — On October 28, Chaoyan Intelligence released its first 35B model, Apex-35B-V1.0. This is a model that balances general capabilities and research capabilities: it can not only complete routine tasks such as Coding, Agent, Q&A and deep search, but also deeply participate in AI research, computational science and robot simulation.

According to the evaluation results released by the company, the model has comprehensively surpassed the Flint model in a large number of evaluations of research capabilities, agents and programming, and reasoning capabilities, reaching the SOTA level. It even outperforms large-size models with parameter scales as high as 500B and 800B on some tasks.

What is more noteworthy than the capability list is its R&D method. It is understood that Chaoyan Intelligence has introduced RSI into model R&D, allowing AI to participate in experimental design and result analysis, and applying the verified effective results to the next round of R&D. In other words, part of the model is "researched" by the AI itself.

The release of Apex-35B-V1.0 is, to a certain extent, a phased verification of Chaoyan Intelligence's RSI roadmap, which also gives a more specific answer to why the company is so sought after by capital.

27-year-old Tsinghua Professor Starts a Business, Letting AI Train AI

A current trend in the primary market is that after the popular track shifts to the AI field, "young prodigies" are obviously more favored by investment institutions, and the average age of unicorn founders in the AI field has dropped sharply from 40 in 2020 to 29 in 2024.

Chen Yongchao of Chaoyan Intelligence is, to a certain extent, such a "young prodigy" type founder. He obtained his bachelor's degree from University of Science and Technology of China, his joint-training doctorate from Harvard and MIT, and joined Tsinghua University at the age of 27, becoming the youngest assistant professor in the history of the School of Artificial Intelligence of Tsinghua University.

"Smart" and "extremely strong learning ability" are almost the keywords mentioned by every investor who has come into contact with him. Investor A once shared a detail with me: Chaoyan Intelligence did not hire an FA, and all work from early communication to subsequent equity changes was completed independently by Chen Yongchao and his team members. "During the contact process, we can clearly feel his extremely strong learning ability, especially for knowledge related to equity investment, he learns, understands and digests at a very fast speed."

This point, in fact, can be traced back to his past experience. From his undergraduate period to the first year of his doctorate, Chen Yongchao successively entered five laboratories, with research directions spanning mechanics, metal materials, nuclear fusion and AI for Science; and he published top journal papers as the first author in each direction, and his achievements in the sophomore year were published in Nature Communications.

When talking about the original intention of starting a business, Chen Yongchao said that he had the idea of starting a business as early as two years ago. During his internship at Google DeepMind, he was first exposed to auto research and AI scientist, which means allowing AI to independently explore new knowledge and then use the research results to improve the model itself. This experience directly determined the direction of his subsequent entrepreneurship. The release of Claude Opus 4.6 in February this year made him feel that the timing had finally come.

"When Opus 4.6 was released, I had a feeling: the existing models have reached a critical point. Although they can make some valuable research, their creativity is limited due to the too heavy blueprint, so we need to train the next-generation model," Chen Yongchao said. "RSI is at the best time point now. Considering the ecological position, I have to start my business immediately."

So in June this year, he gave up multiple overseas offers including those from DeepMind, returned to China and founded Chaoyan Intelligence.

It should be pointed out that RSI is not a newly emerging concept. As early as 1965, mathematician I.J. Good put forward the idea of "intelligence explosion", but it was last year that the industry really picked it up again. In May, Google DeepMind's AlphaEvolve was able to use Gemini to generate candidate algorithms and automatically select the best ones.

Entering 2026, all parties have taken more intensive actions: Anthropic released "When AI Builds Itself", disclosing that more than 80% of the code merged into the production codebase was written by Claude; OpenAI set up a special RSI team; Meta FAIR released AIRA-2. In China, MiniMax and Tencent Hunyuan have successively launched self-evolution systems, and Zhipu has also written "Recursive Self-Improvement" into its technology roadmap for the first time in its recent financing announcement of about 5 billion US dollars. At the same time, a large number of start-up teams in the RSI direction have emerged in both Silicon Valley and China.

Although all are focusing on RSI, their entry points are different. Many companies in Silicon Valley start from the Agent layer to build a system that can automatically conduct scientific research, while Chaoyan Intelligence chooses to build a Recursive Intelligence system at the basic model level. According to the company, this capability is expected to cover AI, mathematics, physics, chemistry, biology, chip design, robotics, industrial manufacturing and other fields in the future.

Now, four months after its establishment, the Chaoyan Intelligence team has nearly 30 members, and the core members come from Tsinghua University, Peking University, Harvard University, MIT and other universities, as well as leading large model manufacturers such as ByteDance, Kimi and Zhipu.

From Apex Research to Apex-35B-V1.0: AI Begins to Participate in "Building Models"

Shortly after its establishment, Chaoyan Intelligence launched its first self-evolution AI system, Apex Research. This system can independently complete the whole process from literature research, method design, code writing, experimental analysis to paper writing and self-iteration with less manual intervention.

The reason why it chooses to cut in from R&D is that in Chen Yongchao's view, the creativity pursued by self-evolution models naturally matches the high fault tolerance of R&D, and "research is not an application scenario, but the engine of RSI itself."

"What the current large models are doing is to compress and remember the existing knowledge of the world, pursuing the ability to be stable. The self-evolution model, on the other hand, creates unknown knowledge of the world. R&D is highly fault-tolerant — 'let it try 10 times, as long as one time is right'. Stability and creativity often cannot be achieved at the same time. A model with a completely safe and sound solution is useless to researchers," he explained.

It is understood that in May this year, Apex Research independently produced 34 machine learning papers and submitted them to ACL Rolling Review for double-blind review. The results came out in July: the average review score was 2.60, close to 2.63 for human submissions, and 2 of them reached 3.67 points — this result has surpassed about 99% of human submissions.

Chen Yongchao mentioned that the team has now built about 100 research projects around multiple fields, which are used on the one hand to evaluate the model's capabilities, and on the other hand to precipitate research trajectories that can be used for post-training, and explore "data self-evolution" to let AI participate in generating research tasks, data sets and trajectories, which are then used for post-training.

It is worth mentioning that during the R&D process of the newly released Apex-35B-V1.0, Apex Research also participated in multiple links including self-evolution training data, automated experiments, training schemes, GPU kernel and inference optimization.

Specifically, in the data link, Apex Research will identify the capability weaknesses of the model and automatically generate targeted training data; in the experimental link, it will analyze the evaluation results and agent operation trajectories, find out the systematic failure modes and optimize the training strategies. For example, in view of the problem that the performance of the agent decreases due to the over-long reasoning context and unbalanced data ratio, the system explored different sequence length distributions, the sampling ratio of short trajectories and long trajectories, and adjusted the composition of training batches according to the number of tokens, to find a more effective training configuration, which improved the reliability and training efficiency of the agent.

In addition, Apex Research extends automated optimization to the underlying computing system. On the training side, the system optimizes GPU load balancing, overlapping execution of communication and computing, dynamic batching, and allocates computing load according to the number of tokens, so as to reduce synchronization overhead and GPU idle time, and improve hardware utilization. Data provided by the company shows that this series of optimizations has increased the training throughput by 45%-60%.

On the inference side, the system optimizes the attention kernel, KV Cache management (i.e. the scheduling of historical key-value cache during inference), and the execution mode of the pre-filling and decoding stages. Under similar load conditions, the decoding speed is increased to 1.6-1.8 times of the original, and the first token latency is reduced by 30%-35%.

According to Chen Yongchao, these improvements run through data, training and infrastructure, and gradually accumulate into a continuously feedback research cycle: the results of each round of research provide the basis for the next round of improvement. In the demo displayed by Chaoyan Intelligence, Apex-35B-V1.0 has been able to participate in tasks such as AI for AI, computational fluid dynamics and robot simulation.

However, in his vision, Apex-35B-V1.0 is just the beginning.

When talking about the long-term goal, Chen Yongchao said that the self-evolution basic large model that Chaoyan Intelligence wants to build will have overall scientific research capabilities equivalent to and gradually surpass 100,000 top AI Scientists covering various disciplines, forming a scalable "AI Researcher Network".

At the end of our conversation, I asked Chen Yongchao a question: Now everyone is talking about AI panic. If one day the AI you created replaces you, will you feel anxious?

Chen Yongchao smiled and said: "If it replaces me, I will applaud it, which means my work has been successfully completed."

Haidian, Once Again at the Forefront of AI

The basic situation of Chaoyan Intelligence's angel round and angel+ round financing has been introduced above. Judging only from the list of investors, this is already a start-up company with the potential to be a "dark horse". However, regarding the popularity of this financing, there are several hidden details worthy of attention:

First, the company took less than one month from starting the financing to finalizing the final equity structure; it originally planned to open one round, but later added consecutive rounds because the quota was not enough to allocate;

Second, the three leading investment institutions of the angel round all sent their core decision-making layers to participate in the communication at the first meeting;

Third, there are no fewer local state-owned capitals competing for the investment quota, but affected by the allocation requirements of capital types, the "large funds" from Beijing, Shanghai and Guangzhou finally appear in the shareholder list, and the company chooses to set up its registered address in the AI Origin Community in Haidian, Beijing.

When talking about the reason for finally choosing to locate the company in Haidian, Chen Yongchao mentioned several key points:

First of all, it is related to his another identity — the assistant professor of Tsinghua University. He said that the reason why he chose to teach at Tsinghua University at that time was precisely because he valued Tsinghua's dominant position in the field of artificial intelligence and its good industry-university-research ecosystem. "You can see that in this wave of AI, whether it is large models or embodied intelligence, many of the best star companies are founded by Tsinghua teachers. If we locate the company in Tsinghua and Haidian, the probability of becoming a leading player will be much higher."

The second is the talent advantage. Data shows that Beijing gathers about 43% of the top AI talents in China, with a total of about 12,300 people, of whom more than 80% are located in Haidian. For a start-up, such a high talent density means lower recruitment costs and faster team building speed.

"If I want to make an appointment with a researcher from a certain model company to chat, since everyone is in Haidian, we can just find a place to have a meal, which is very difficult to achieve in other places," Chen Yongchao said.

In addition, he also specially emphasized that in the process of cooperation with Zhongguancun Science City, in addition to capital, the other party can give full support and coordination on various supporting resources required for enterprise development. Moreover, behind this support is the whole "Five Parties and Six Forces" mechanism of Haidian as the backing.

The so-called "Five Parties and Six Forces" is a set of achievement transformation mechanism pioneered by Haidian to accelerate original innovation and break through the blocking points and difficulties from the laboratory to the production line. Based on this mechanism, Haidian has specially launched the "One School, One Policy" working model for universities and research institutes, forming working groups around key universities to carry out precise services. The Tsinghua Special Working Group is one of the 9 university working groups that have been fully operational.

For the Chaoyan Intelligence project, the Tsinghua Special Working Group has also done a lot of pre-service work, such as the coordination of office space, which is exactly what it contributed to.

"It can be seen that Haidian is truly systematically incubating cutting-edge technology enterprises and building a complete science and innovation ecosystem. I just returned to China to start a business and started here. The whole process made me feel not stylized policy docking, but real warmth, which gave us great confidence," Chen Yongchao said.

Abundant talent reserves, strong scientific research resources, and multi-dimensional empowerment from the government level together constitute the reasons why Chaoyan Intelligence takes root here.

In Haidian, there are many similar stories. Within this 430 square kilometer area, more than 2,000 AI enterprises have been gathered, covering the whole industrial chain from computing power, algorithm models to industry applications.

More critically, the prosperity of this "innovation ecosystem" lies not only in the number of enterprises, but also in the output quality. Zhipu AI, Starbot Era, MemryX, Baichuan Intelligence, Shengshu Technology, CoreVision Technology... a number of benchmark enterprises in their respective fields have all been born and grown here.

To sum up, it is the complete science and innovation ecosystem that makes countless excellent entrepreneurs choose Haidian at the beginning of their business, and it is also such profound accumulation that enables Haidian to seize the opportunity in cutting-edge trends like RSI again and again.

It is foreseeable that such a positive cycle will continue to occur on this land.

This article is from the WeChat Official Account "ChinaVenture", written by Wang Manhua, authorized for release by 36Kr.