Bonus Exclusive | Baichuan Intelligence Hits the Brakes, Adjusts Medical ToB, and Basic R&D Comes to a Halt

36Kr's "Workplace Bonus" (ID: ZhiChangHongLi)

DeepSeek's sudden popularity has continued the myth of the AI industry, but Baichuan Intelligence, which benchmarked against OpenAI during its financing, has not welcomed spring.

"Workplace Bonus" has exclusively learned from multiple independent sources that the following changes have recently occurred at Baichuan Intelligence:

1. The training of the basic model has stopped, and the relevant core person in charge has proposed to leave and start a business, focusing on the AI Coding direction.

2. The person in charge of the medical product line is preparing to leave and start a business.

3. The human resources department is making relevant adjustment and optimization moves, and the number of personnel may reach a scale of one hundred. The medical To B business, which was part of Baichuan's key business before, may become the key target for layoffs.

"Workplace Bonus" verified the above information with Baichuan Intelligence. The public relations department denied the above statements and responded that "none of them are true. Baichuan has always regarded AI medical as a development direction and will always work on the medical enhanced large model."

Multiple sources told "Workplace Bonus" that this adjustment involves nearly a hundred people, including those in technology and the medical To B business. In addition, Baichuan Intelligence stopped pre - training last mid - year. "They returned the pre - training computing cards, and almost all the people in the pre - training group have left." In fact, this is not the first time Baichuan has laid off employees.

Earlier today, LanJing News reported that Jiao Ke, the co - founder of Baichuan Intelligence, has left, and another co - founder, Chen Weipeng, will also leave. He is currently going through the internal process. This is also consistent with what the previous sources said.

According to "Intelligent Emergence", Baichuan Intelligence's business service department responsible for the To B business was laid off before, and all employees signed the separation agreements on March 3. "Workplace Bonus" learned from an informed person that about 40 people in the B - end group were laid off in the previous adjustment and optimization.

With the suspension of model R & D, the departure of core personnel, and the layoff of the previously "focused" medical To B business, Baichuan Intelligence has pressed the pause button in the accelerating industry.

Suspension of basic model training, core person in charge applies for resignation ╱ 01

Fear before the arrival of R2, "Cutting off the arm" of the medical To B? ╱ 02

Emergency brake, seeking opportunities in To C ╱ 03

Suspension of basic model training, core person in charge applies for resignation

At the beginning of this year, the departure application of the person in charge of model R & D technology was also an important part that prompted Wang Xiaochuan to make the decision of adjustment and optimization. "Workplace Bonus" learned from a senior headhunter that the industry still has a high recognition of Baichuan's basic model team.

If the core technology person in charge leaves, a new technology person in charge needs to be recruited later, and the R & D work can only be promoted after they get used to the work. This series of processes not only requires a lot of money and energy, but time is also a more important cost.

But Baichuan Intelligence can't wait that long. An industry insider told "Workplace Bonus" that the current situation is caused because Wang Xiaochuan insisted on shifting to the medical field last mid - year, resulting in the lag of basic technology.

Shifting to the medical field means allocating more computing cards and computing power to train new models. At that time, there was also the training work of the large financial model, so there was not much left for the basic model. Also last mid - year, Baichuan's pre - training was almost suspended. "They hardly did it anymore, and the newly ordered computing cards were all returned."

When Baichuan was founded, it aimed to be the "Chinese ChatGPT" and received a lot of financing. At that time, Wang Xiaochuan shouted the development strategy of "super model + super application" driven by two wheels. In 2023, Baichuan Intelligence accelerated its pursuit. In just half a year, Baichuan Intelligence released its open - source model baichuan - 7B. The model's ability once "surpassed" ChatGLM and "outperformed" LLaMA. And 26 days after the release of the open - source 7B model, Baichuan Intelligence open - sourced its commercial large model Baichuan - 13B.

In the first half of 2024, Wang Xiaochuan frequently mentioned in public that "medical is the 'pearl on the crown' of large models." The medical field had long planted a deep seed in Wang Xiaochuan's heart. In April 2023, in the first conversation between "Intelligent Emergence" and Wang Xiaochuan, Wang Xiaochuan did not hesitate to express his love for life and health and his interest in the medical field.

However, Wang Xiaochuan also said in a conversation with "LatePost" that "it's difficult to get financing without telling the story of the general model."

Although slowly, Baichuan's basic model was still being developed last year. "At least at that time, multimodal and reinforcement learning were still being worked on," "Workplace Bonus" learned from an insider. "But during the Spring Festival this year, when DeepSeek came out, Wang Xiaochuan suddenly didn't know the meaning of doing model research."

DeepSeek is an open - source model. On the other side of technological equality, the value of many closed - source self - developed models in the To B market has been weakened, including Baichuan's. Employees of other large model companies revealed to "Workplace Bonus" that some B - end users directly proposed whether they could access the DeepSeek model while deploying closed - source models. It's really ridiculous.

And the disagreement on the technical path with the person in charge of model R & D made Wang Xiaochuan even more confused about the future. An informed person told "Workplace Bonus" that this is also an important reason for the departure of the technology person in charge. Wang Xiaochuan previously thought that not much cost needed to be invested in basic model training, and he judged that his own technology couldn't catch up with DeepSeek within half a year. So after the Spring Festival, he basically did nothing. But the former didn't agree with this approach.

Fear before the arrival of R2, "Cutting off the arm" of the medical To B?

Last year, Baichuan was frequently reported with financing news, and its technology and business were also progressing. The company as a whole was still in the limelight.

An informed person said that earlier last year, Baichuan planned to layout in the financial field. Entering the financial field is because the financial industry has a more perfect marketization, high returns. And the financial industry has relatively abundant funds, is willing and more capable of exploring artificial intelligence.

Judging from the public information of "Baichuan Intelligence Business Services", Baichuan's layout in the financial field is quite deep. In 2024, Baichuan Intelligence's business service department brought many leading enterprise customers and partners to Baichuan, such as Bank of China, Beidian Digital Intelligence, China Merchants Bank, Sunyard, and Tiankai Group.

The To B business is the heart of Baichuan's self - financing. In 2024, it brought nearly 100 million yuan in revenue to Baichuan. Baichuan also previously expected that the To B business would increase several times in 2025. The growth of the To B business gave Baichuan great confidence. Ru Liyun, the co - founder of Baichuan Intelligence, once mentioned in an exclusive interview with "Caixin" that the company's order signing amount in 2025 would be at the level of 1 - 2 billion yuan.

But everything changed after DeepSeek came out.

Not only Wang Xiaochuan, but also the six major large - model companies are all thinking: A stronger, more popular, and open - source model has come. What's the meaning of continuing our large - model R & D? What if the technology is updated again after we've spent a lot of time and energy to catch up? The customization cost of the To B commercialization model remains high. Should we continue?

The frequent news of the early release of DeepSeek - R2 has further aggravated the concerns of the six major large - model companies. On February 25, Reuters reported that DeepSeek is accelerating the launch of its next - generation AI model R2. It was originally planned to be released in May, but the company is working hard to launch it as soon as possible.

Another person close to Baichuan Intelligence told "Workplace Bonus" that the medical commercialization business that Baichuan Intelligence is currently doing is not profitable. The high customization degree of large - model projects, which leads to high costs and other problems, is also the key reason why the industry has always been pessimistic about the To B business model of large models.

An expert from Peking Union Medical College Hospital who has been involved in large - model research also told "Workplace Bonus" that medical dedicated models have poor universality. Due to the differences in clinical experience, the medical dedicated models trained by hospitals with the data from their own research centers are only applicable to their own centers. There will be obvious errors when applied to other hospitals.

Moreover, the amount of data used by a single hospital to train large models is very limited. And due to the differences in clinical levels between hospitals, it is difficult to interoperate data. In addition, out of the protection of patients' privacy, hospital data is prohibited from leaving the hospital. This also increases the difficulty of training medical models.

The medical To B business is originally a hard nut to crack. In terms of commercialization, in China's system, the medical industry is one of the industries that cannot be a profit - making institution. Although hospitals embrace AI with a positive and open attitude, the annual expenditure of millions of yuan for customized models is still a huge sum for hospitals.

The influx of large companies into the medical field has made it even more difficult for Baichuan. At the beginning of the year, Ant Group and Haodf.com officially announced the completion of the acquisition; recently, Quark recruited a person in charge related to the medical large model; on March 7, Huawei officially announced the establishment of a healthcare business group, focusing on AI - assisted diagnosis and medical large - model applications.

In this "critical" situation, Baichuan "cut off the arm" of the medical To B business.

Emergency brake, seeking opportunities in To C

The industry is quite sighing about Wang Xiaochuan's decision to adjust the business service department. "Compared with adjusting the financial business, people would rather see him adjust the medical business," an investor said.

Because the financial industry is originally a natural big spender in IT. Information is digitalized, and the amount of data is also large, which is more suitable for achieving some results in the B - end. Baichuan itself is not weak in technology. In theory, it can still achieve considerable returns by deeply exploring the To B scenario and accumulating its own know - how.

The medical industry is originally not very suitable for B - end commercialization, and the know - how barrier is very high, the growth is relatively slow, and the PS multiple is low [1].

This decision also reflects Wang Xiaochuan's lack of optimism about the To B business model. But many practitioners point out that Wang Xiaochuan's understanding of the To B business is indeed relatively shallow. The six major large - model companies still have the opportunity to build advantages in To B commercialization, because in To B commercialization, in fact, it depends on who can better meet the needs of customers, and the model ability is secondary.

"And it can't all be blamed on Wang Xiaochuan. The logics of To B and To C are indeed quite different, and it's common that they are not interoperable," an industry insider said.

According to "Intelligent Emergence", Lingyi Wanwu recently split its digital human business. Before that, it "packaged" the pre - training and Infra teams and "generated self - financing" through splitting. An industry practitioner told "Workplace Bonus" that Lingyi Wanwu had to do this because of the urgent shortage of funds on its account.

But different from Lingyi Wanwu, Baichuan still has money on its account. An industry insider close to Baichuan estimated that "there should still be more than 3 billion yuan on Baichuan Intelligence's account." What Wang Xiaochuan is doing now may be to first streamline the team cost, then make long - term plans and wait for opportunities.

On March 17, "Intelligent Emergence" also mentioned that Baichuan is discussing how to combine the medical business with "Baixiaoying" to develop C - end medical products.

The recent popularity of Manus has verified the C - end's enthusiasm for large models and agents. But at present, there are only a few companies in the market that can create distinctive agent products. With Baichuan Intelligence's technical background, it is still capable of doing so, which also creates a good development opportunity for Baichuan. And for Wang Xiaochuan, it may not be a bad thing to return to the field he is good at.

As long as Baichuan stays in the game, there is still a chance.

[1] The PS multiple (Price - to - Sales Ratio, P/S) is a financial indicator used for company valuation. It measures the market's valuation of a company's sales revenue by comparing the company's total market value or stock price with its sales revenue. A high PS multiple indicates that the market has high expectations for the company's future growth potential and profitability; a low PS multiple may indicate that the company is currently undervalued by the market, especially if the company has stable sales revenue and potential growth opportunities.

Written by | Chen Jiahui

Edited by | Chen Tong

Typeset by | Song Jiuqing

Cover image | Marcelino (Unsplash)

36Kr's "Workplace Bonus" (ID: ZhiChangHongLi)

36Kr's "BonusBonus" (ID: ZhiChangHongLi)

This article is from the WeChat public account "BonusBonus" (ID: ZhiChangHongLi), author: Chen Jiahui, published by 36Kr.