The "godfather" of online education ventures into AI education: Raised 1 billion yuan in three years with monthly revenues exceeding 10 million yuan

Text by | Zhou Xinyu

Edited by | Su Jianxun

When the World Digital Education Conference defined 2025 as the "Year of Smart Education", a Chinese AI education company founded two and a half years ago has already become a unicorn.

Liu Wei, the founder and COO of "Dance with Love", revealed to media such as "Intelligent Emergence" that in the two and a half years since its establishment, "Dance with Love" has completed four rounds of financing, with a total financing amount of approximately $150 million. Currently, the company's valuation is close to $1 billion.

△ Shareholders of "Dance with Love". Photo source: Taken by the author

Zhang Huaiting, the CEO of "Dance with Love", and Liu Wei, the COO, are a group of practitioners with particularly keen senses.

The former was once the core person in charge of Baidu's commercialization system "Fengchao" and a co-founder of Gaotu (GenWhoXue). The latter was once the vice president of Gaotu Group and the general manager of Gaotu Classroom, and is the core inventor of the "Two-Tutor Live Large-Class Course".

Some shareholders once told us that almost at the same time as the release of ChatGPT, Zhang Huaiting already had the idea of reinventing the education industry with AI. In May 2023, in the same period as the "Six Little Tigers of Large Models", Zhang Huaiting and Liu Wei, together with four technology and education practitioners, jointly founded the AI education application company "Dance with Love".

Since its establishment, "Dance with Love" has always been one of the hottest targets in the primary market. It has received seed-round investments from first-tier institutions such as Gao Rong Capital, Sequoia Capital, ZhenFund, Qifu Capital, BlueRun Ventures, and Xianfeng Changqing. After the investment, its valuation reached $125 million.

According to "Intelligent Emergence", one year after its establishment, the valuation of "Dance with Love" soared to $800 million.

The enthusiasm of capital stems from the rapid progress in implementation.

At a live recruitment fair, Wang Lin, the co-founder and technology leader of "Dance with Love", revealed that the company's revenue in the month of its establishment was only a few hundred thousand yuan, and it reached over one million yuan the next month. Now it has tens of millions of yuan every month.

But compared with the bustling money stories, the founding team and the business itself of "Dance with Love" are very low-key. What most people are curious about is: What exactly has "Dance with Love" done in the two years since its establishment?

On November 19, 2025, at the headquarters of "Dance with Love", Liu Wei, the founder and COO, talked about the progress in the two and a half years since the start of the business for the first time.

△ Liu Wei, the founder and COO of "Dance with Love". Photo source: Taken by the author

This veteran in the education and training industry still retains the habit of teaching on stage before he steps onto the stage. For example, every five minutes, he will throw a question to the audience to attract their attention; when he comes to the key points, he will use exaggerated gestures to emphasize the content on the PPT.

The encouraging slogan culture of Gaotu has also been carried forward. For example, on the desks of the user growth department of "Dance with Love", there is a slogan pasted: Some people are suitable for sprinting, while others are suitable for long-distance running.

But in the two and a half years since the start of the business, what Liu Wei thinks more about is the difference between AI education and online education, and what kind of product innovation AI education needs.

"Agriculture", this is Liu Wei's metaphor for online education. He told "Intelligent Emergence" that the logic of product development in online education is "plant as many fields as there are people": on average, one teacher in online education teaches 100 students, so one million students need 10,000 teachers.

Therefore, in Liu Wei's view, in online education, it's all about organizational ability. How to recruit more teachers and expand the scale? How to attract stronger teaching staff at the lowest cost? How to reduce the variance in teachers' levels? These are the questions he had to think about every day at Gaotu.

But AI education is a field where it's impossible to build a moat with human resources - the competition has entered a completely different field: AI technology.

Liu Wei described that the biggest difference between "Dance with Love" and Gaotu lies in the former is essentially a technology company, while the latter is an education company - in the past at Gaotu, when facing business problems, the team would think about how to pile up human resources and compete for teaching staff; but now, "the first thing that comes to mind is how to use technology and products to solve the problems".

Polishing AI technology is a completely different process from the scale race in online education, more like brewing tea slowly. In 2017, Gaotu had only a seven - person team and a revenue of only 20 million yuan. However, in 2018, Liu Wei increased the revenue to 200 million yuan, 2 billion yuan in 2019, and 6.4 billion yuan in 2020.

But in the first two years of "Dance with Love", the team's focus was not on the commercialization of AI products, but on building a technology system around AI digital humans, real - time audio - video conversational AI education models, and intelligent interactive voice models.

△ The real - time audio - video conversational AI education model independently developed by "Dance with Love" based on an open - source model. Photo source: Taken by the author

For example, in 2023, the voice delay of AI models was still a problem that troubled many enterprises. It wasn't until the release of Doubao in August 2023 that the delay in voice interaction was significantly reduced.

At that time, "Dance with Love" quickly assembled a voice system R & D team - later, the real - scene recognition accuracy of this voice technology reached 95%, and the response speed reached the millisecond level.

△ The self - developed voice technology of "Dance with Love". Photo source: Taken by the author

Liu Wei told "Intelligent Emergence" that once the underlying technology is refined to solve business problems, indicators such as revenue and growth will not be a problem.

Another change in the team lies in the product.

"Online education thrived on the live large - class courses and also declined because of them." Liu Wei reflected.

The "Two - Tutor Live Large - Class Course" was once an important source of revenue for online education. However, he found that the penetration rate of this model was not high in first - and second - tier cities, and offline class courses were still the mainstream. "In terms of teaching effect and willingness to pay, one - on - one individualized teaching is definitely the best."

In his view, the generalization ability and self - learning mechanism of large AI models can precisely make one - on - one customized teaching possible.

For this reason, in May 2025, two years after its establishment, "Dance with Love" launched the product "Love Learning", which features a one - on - one AI tutor for all subjects and users of all ages.

△ The process of "Love Learning" tutoring math problems. Photo source: The official website of "Love Learning"

Before the launch of "Love Learning", the doubts Liu Wei often faced were: Can users accept AI tutors? Are they willing to pay for AI tutors? And, are AI tutors really effective?

He told "Intelligent Emergence" that before its launch, "Love Learning" also went through several months of internal testing. Some of the tests were: not telling users in advance that it was an AI tutor and checking the course completion rate; then informing them and checking the refund rate.

The test results showed that the course completion rate of "Love Learning" even exceeded that of the live large - class courses; at the same time, almost no one withdrew from the course because of the AI.

Of course, when the wave of large AI models arrives, the opportunities in AI education won't belong only to entrepreneurs.

Since 2024, established online education institutions such as Yuanfudao, Zuoyebang, and Youdao have all combined their businesses with AI and launched software and hardware products such as AI problem - solving, AI oral English practice, and AI learning machines.

In the competition with new and old players, as a startup with limited resources, the most important thing is to make choices and concentrate resources to build a moat.

Liu Wei's experience is to first enter the fields that are in urgent need of users and are more difficult.

For example, in the early days of the startup, among the three directions of digital human teachers - 2D cartoon, 3D cartoon, and real - life human - the team finally chose real - life human. At that time, "real - life human was the most difficult, and most third - party platforms couldn't support it at that time," Liu Wei said, but the acceptance of real - life human teaching was the highest.

"AI is developing faster than I expected," Liu Wei responded to the media. Even though he can't predict the year when the commercialization of AI education will explode, it's destined to be a fast - paced race. "When we started the business in 2023, I thought it would take three to five years for an AI teacher to complete a course, but in fact, we achieved it in one or two years."

The following is a compilation of the group interview with "Dance with Love" by "Intelligent Emergence", with the content slightly edited:

Promising a specific score increase is an over - commitment

Media: Will the accuracy of AI be the same in different subject teaching scenarios? For example, will it be more accurate in science subjects than in liberal arts subjects?

Liu Wei: To avoid "hallucinations", we can't directly use large models; we have to use our own trained models. We have a large number of internal evaluations to ensure relatively stable output.

People think that liberal arts have no rules, but actually they do. Both liberal arts and science have their own rules, and they are all rational. As long as there are rules, AI can learn.

Media: Gaotu proposed the "Two - Tutor" model, where a famous teacher teaches a large - class course, and then ordinary teachers assist students in understanding. Do AI teachers also need real - life teachers to assist?

Liu Wei: In the whole - process interaction, how should AI and humans divide the work? How can they help children learn better more smoothly? Exactly how much work should AI and humans do respectively? What is their connection?

We are still in the process of exploring this whole process.

Media: Specifically regarding teaching results, how much can "Love Learning" improve students' grades?

Liu Wei: Essentially, no matter how excellent a teacher is, the interaction between the teacher and the child is two - way, and the child is the main body, not the teacher. No teacher can guarantee that a child will definitely get 100 points.

But some institutions market in this way, which is not a good phenomenon. I'm not saying that I don't care about children's grades, but promising a specific score increase is irresponsible.

The biggest guarantee of "Love Learning" is one - on - one teaching, and its effect and efficiency are naturally much higher than those of class courses.

Media: "Much higher" is too vague. Users need data to believe it.

Liu Wei: We get data for each question. For example, for a certain question, a child's initial correct answer rate was about 50%. After being guided by AI, the correct answer rate increased to 80%.

Media: The increase rate of the correct answer rate is a very Internet - thinking dimension. But to convince parents to pay, you need to show a real score - increasing effect.

Liu Wei: We can't promote it like that, or it will be an over - commitment.

For example, if one exam is easy and the next is difficult, it's hard to reflect the effect by comparing the scores of the two exams. So you'll find that no institution promotes in this way.

If we focus on what the child has learned, how well they have mastered it, and what the related experience is, we can provide data. But we can't say exactly how much the score will increase.

Media: Once the AI has "hallucinations" or can't solve a problem, will there be a real - life teacher to back it up?

Liu Wei: We focus a lot on the training stage. We have a large number of tests and data in the training process, and the trainers themselves will score these data and results.

So first, we put our efforts into the training link to ensure that the process is controllable and the knowledge imparted is correct. Secondly, there is also a data flywheel in the interaction process with the child.

I can't say that our model will never make mistakes, but the probability is very low because a large amount of data has been pre - trained.

Media: "Love Learning" focuses on the teaching scenario. What other implementation scenarios are there for AI education?

Liu Wei: Educational products can be divided into two major categories. One is the solution - oriented type, such as "Love Learning" and Gaotu's large - class courses. The other is the tool - oriented type, such as Xiaoyuan's question - searching and Zuoyebang's photo - searching.

The problems they solve are different. Tool - oriented products rely heavily on students' initiative, while solutions don't rely on students' initiative. They have a systematic process to help children achieve their goals.

If you want to rely on self - learning, you don't necessarily have to sign up for a class. Many students with strong self - learning abilities can prepare for the postgraduate entrance examination on their own. However, younger children may have weaker self - learning abilities, so K12