36Kr Exclusive | The Head of Generative Large Models at Huawei Noah's Ark Lab Launched a Home Embodied Intelligence Startup, Raising More Than 1 Billion Yuan Within One Year

Source / Enterprise

This article is about 6900 words, with a recommended reading time of 17 minutes

Author | Ou Xue

Editor | Yuan Silai

Among the embodied AI startups spun out from Huawei's ecosystem, Noein Intelligence is one of the companies with relatively smooth financing progress.

Hard Krypton has learned that this company, founded in August 2025, recently announced the completion of a single angel +++ round financing of several hundred million RMB. This round is led by a JD-related fund, with participation from Zhengxin Gu Capital, Nanshan Strategic Emerging Industry Investment, and Walden International. Only over one month has passed since its previous 500 million RMB financing round.

Before founding Noein Intelligence, Li Yinchuan had R&D and implementation experience covering embodied AI, large models, and autonomous driving. At the age of 25, he graduated from the joint doctoral program of Beijing Institute of Technology and Columbia University and joined Huawei. At 27, he served as the head of the generative large model direction at Huawei's Noah's Ark Laboratory. At the age of 30, he left Huawei to establish Noein Intelligence. In his view, home robots are a critical opportunity to translate large model capabilities into tangible products.

After the 2025 Spring Festival Gala, embodied AI has garnered widespread attention from the capital market. Videos of robots folding clothes, pouring coffee, and tidying tables have frequently appeared in public view, showing the public the possibility of robots entering households. However, moving from demonstration scenarios to daily use requires handling far more complex situations: if the room is changed, the set of cups and plates is replaced, or the type of housework is altered, can the robot still complete the task stably?

This is exactly the problem Noein Intelligence aims to solve. Li Yinchuan recalled that in the early days of his entrepreneurship, the industry focused more on industrial scenarios and imitation learning, while he chose a technical route centered on generalization capabilities and oriented toward home scenarios. In his view, for robots to enter households, two problems must be solved simultaneously: adaptability and cost. The model needs to handle tasks in different households, and the product price must be within the acceptable range of ordinary consumers.

To this end, they developed the technical base GLOW, a generative embodied large model architecture. It is coordinated by multiple modules, enabling robots to gradually evolve from "reproducing trained actions" to "solving tasks that have never been fully encountered before". In terms of data, 90% of the data Noein uses is synthetic data, which can reduce the overall cost by 10 to 100 times. In terms of training, the world model can deduce the consequences of actions without generating complete video frames. In terms of execution, a single inference of the model can output a complete action sequence, reducing invocation frequency.

Meanwhile, Noein Intelligence launched its first consumer-grade home embodied robot, KNOWIN-X1. It adopts a fully foldable dual-arm configuration, with a height controlled within 40 cm after folding, and can adapt to home spaces such as desktops, floors, sofas, and washing machines.

KNOWIN-X1 (Source / Enterprise)

Of course, these products are still quite far from large-scale implementation. The continuous capital injection is partly due to Noein Intelligence's high talent density. The company has expanded to more than 200 employees in less than a year. Among them, over two-thirds of the algorithm team hold a doctoral degree or above, and half of the members graduated from top 30 institutions in the QS World University Rankings. The team has gathered multiple "Genius Youths" from Huawei.

Li Yinchuan remains optimistic about the industry's prospects. He believes that the entire track may experience fluctuations in the near future, but it is far from the stage where the bubble completely bursts. The key indicator to watch is when capability emergence occurs in embodied large models.

There is still no consensus on this point in the industry. Some people believe that the robotics industry will not have a node similar to the "ChatGPT moment", but Li Yinchuan argues that intelligent emergence is only a matter of time.

"If embodied models one day truly achieve capability emergence, the commercial value of robots will undoubtedly far exceed that of large language models, and I think it may be more than 10 times that," he told Hard Krypton.

We talked to Li Yinchuan about his reflections since starting his business and his observations on home embodied AI. Below is the conversation between Hard Krypton and Li Yinchuan, founder of Noein Intelligence.

Starting a Business After Leaving Huawei

Hard Krypton: What was on your mind when you decided to leave Huawei Noah's Ark Lab to start your own business? What was the decisive factor?

Li Yinchuan: The main reason I left to start a business is that I wanted to build the most generalized embodied large model in the world and truly incubate its commercial value, so I chose to focus on To C embodied AI. When I was at Huawei Noah's Ark Lab, I felt that if I went back to work on To B and To G scenarios, the support for R&D of embodied large models would be relatively limited. Huawei's strategy is to make Ascend a solid "black soil" platform for everyone to use, rather than developing all upper-layer applications on its own. That's why I was more eager to go out and build home robots.

Hard Krypton: Did you decide on the home robot direction right away back then? What were your considerations?

Li Yinchuan: I was probably one of the first people in China to publicly talk about home robots back then. I had two non-consensus views.

First, in the first half of 2025, the general consensus in the industry was that embodied AI was all about operation, imitation learning, and factory deployment. But I was arguing at that time that we should take a technical route that focuses on generalization, and truly build a generalizable embodied model instead of an overfitted one. Only with generalization capabilities can we have the opportunity to enter the To C home scenario. That was our first non-consensus choice in technical route selection.

Second, after learning about the selling price of robots on the market at that time, we were confident that we could reduce the cost of robots through a new vertical integration approach. Only when the cost comes down can To C scenarios be viable. A robot cannot be priced at hundreds of thousands of RMB, which ordinary consumers simply cannot afford.

With these two non-consensus judgments, I thought this cause had great potential, so I decided to start this business.

Hard Krypton: You said you are confident about cutting costs, what are the specific bases for this judgment?

Li Yinchuan: Take embodied data cost for example. We have calculated that not all data needs to be collected through physical robot operations, and this view has gradually become a consensus in the industry now. 90% of our data is synthetic data, which can reduce the overall cost by 10 to 100 times.

After the data cost is reduced, the next step is the robot body. After full discussions among our core team, we believe that by fully self-developing all components instead of purchasing parts from others for assembly, we can also keep the cost at a relatively low level.

Hard Krypton: You have been emphasizing the generalization capability of the model. But the basic capabilities of current home robots are not particularly strong. Is it too early to emphasize generalization at this stage?

Li Yinchuan: An important reason why the interaction capabilities of current home robots are not strong enough is that they lack generalization, which is exactly why we need to develop generalized models. Many architectures in the industry today are not really moving toward generalization, but are competing to perform better on single-point tasks.

Most embodied companies are releasing cool demos, but they have not given enough consideration to real implementation. Current home robots are somewhat like the early stage of autonomous driving: achieving good performance on a specific section of road makes people feel that the L4 and L5 era is coming soon, but in fact, there is still a long way to go.

For home robots to truly land, the core is to make their basic capabilities sufficiently universal. The hardest part of getting robots into households is not completing a fancy demonstration once, but stably finishing the work in different households. This is the most challenging part, which is why we emphasize building the most generalized large model.

Hard Krypton: Where do you think the current real development level of the embodied AI industry stands?

Li Yinchuan: The entire track is in a good state overall. Embodied large models have not yet reached the moment of capability emergence, but some leading scholars judge that this moment is not far away, with some predicting it will come in one or two years, and others in three to five years.

When capability emergence truly arrives, I believe the commercial value of the robotics track will have far greater potential than large language models. Currently, the valuation of large language model companies is generally an order of magnitude higher than that of embodied companies, with an average difference of about 10 times. Their current profitability is not that high, but the market has already seen the relatively clear growth space after capability emergence.

But if embodied models one day truly achieve capability emergence, the commercial value of robots will undoubtedly far exceed that of large language models, and I think it may be more than 10 times that. So looking back, if embodied models truly achieve emergence, the valuation of embodied companies will still have huge room for growth.

The entire track may experience fluctuations in the near future, but I think it is far from the stage where the bubble completely bursts. The key still depends on when capability emergence occurs in embodied large models.

Hard Krypton: Why do you think the progress of the entire industry is not fast enough now?

Li Yinchuan: I think most embodied companies, and the industry as a whole, are not focused enough. Everyone has been moving along the direction of competing for cool demos, which only pushes market expectations higher and higher, but gets further and further away from real delivery.

If the industry can focus more on truly valuable scenarios, work hard on data to increase data volume and reduce data cost, the development of the entire industry will be much smoother.

Hard Krypton: Is this because everyone feels that real implementation is still far away?

Li Yinchuan: Yes, so this development trend will lead to an unhealthy state of the entire industry.

Hard Krypton: Has the investment logic in the embodied AI market changed now? What are investors paying more and more attention to?

Li Yinchuan: I think changes will happen very soon. Previously, the industry was in the stage of concept verification and high expectations, and capital paid more attention to technical prospects. But as products gradually move to real scenarios, the market logic will naturally return to essence.

Therefore, in the future, investors will no longer simply pay for the team's background, but will pay more attention to whether the enterprise can truly land in scenarios and form a sustainable commercial closed loop and self-sustaining capability.

Hard Krypton: The development of embodied AI is very much like repeating the story of autonomous driving. Do you think it will follow a similar path in the future?

Li Yinchuan: There will definitely be many similarities. There are many autonomous driving companies, some insisting on developing L3, and some aiming for L4 from the very beginning, and obvious differentiation has emerged later. The same will happen to embodied companies.

I am not optimistic about some autonomous driving companies that positioned themselves for L4 or even L5 from the start. I think the companies with the greatest commercial value are those that positioned themselves for L3 in the early stage and moved forward step by step. There are many lessons we can learn from this.

Hard Krypton: How do you classify the intelligence levels of robots?

Li Yinchuan: We previously defined the intelligence capability standards from L1 to L5. L1 is more like toys, remote-controlled and DIY products, with recreational commercial value. L2 can complete basic housework, but only in limited and fixed scenarios, such as completing a single task in a fixed environment. L3 can complete a long chain of task closed loops in one household, where humans can stay on standby to take over, allowing the robot to run automatically for a long time, and achieve a relatively high success rate in tasks such as tidying, cleaning, and laundry, with potential comparable to a basic home assistant. L4 can further match most of the capabilities of a nanny, while L5 means the robot has more general capabilities, and can even complete many tasks that originally require human labor, which is a more distant goal.

But robots do not need to develop to L4 or L5 to generate commercial value. Our priority is to fully unlock the commercial value of L3, while continuously exploring toward L4 to verify the generalization capability of the model. But it is still relatively difficult to incubate the commercial value of L4 prematurely at this stage.

Hard Krypton: What is the main potential of L4?

Li Yinchuan: If we truly achieve L4, the robot needs to fully match most of the capabilities of a nanny.

But in fact, most consumers will not have as high expectations for robots as they do for nannies when buying robots. If the price is not expensive and the robot can stably help with some housework, it can already generate great value.

For example, the expectations for sweeping robots are very clear: it only needs to clean the floor well, but it has already formed a huge market. If robots can do a good job in many housework tasks, their commercial value can already be fully incubated.

Meanwhile, the capabilities formed in home scenarios can also be extended to some semi-To C scenarios, such as greeting guests, handing water, providing companionship, and other commercial scenarios, without necessarily fully reaching the capabilities of a nanny.

But if you want the robot to tidy up the house like a nanny, the requirements for speed, success rate, and even capabilities such as cooking will be much higher.

Hard Krypton: Would it be more realistic to develop dedicated capabilities for specific scenarios first, and then gradually move toward generalization?

Li Yinchuan: From the perspective of technical routes, developing dedicated capabilities actually provides little help for building generalization capabilities, and the technical differences between the two are very large.

Some companies may develop dedicated models first, validate the commercial path, make profits, and then incubate generalized models. This is of course understandable from a commercial perspective: do easy things to make money first, then tackle complex things. But from the perspective of technical route evolution, I don't think it is a natural path.

Building an overfitted system and building a large model are essentially two different things. Language models could write poetry ten years ago, and even a tiny LSTM can write ancient poetry and recite texts, but it is not until today's large models that huge commercial value is generated. The reason is that they have stronger reasoning capabilities, can understand human intentions, and are sufficiently generalized, so that productivity is truly released.

The same applies to embodied AI. If everyone keeps working hard on overfitted scenarios and puts all engineers into the engineering delivery of single-point tasks, these capabilities will not naturally precipitate into the general model base. From the perspective of technical routes, it provides little help for the final generalized model, and can only generate some revenue in advance.

Hard Krypton: How does Noein solve the problems of commercialization and revenue while adhering to the generalization route?

Li Yinchuan: This is also a problem we are continuously solving. We have found some methods to continuously tap commercial value along the way while adhering to the long-term goal of generalized large