Opening Remarks
Liang Wenfeng
Welcome, everyone.
When we first started this company, we were not thinking about how much money we would eventually make, whether we would go to the capital markets, whether we would list, or anything like that. That was never the original intention. The first few dozen people who joined us did not think that way at all. If they had, they would not have come.
Broadly speaking, we began this work with a great deal of goodwill toward the world. We believed this was useful to humanity. It was something beyond money. Of course, once the potential value became enormous, other temptations appeared. That is a separate matter. But our original intention, our vision, and the vision we have kept to this day have never been built around maximizing commercial profit. I think that point is crucial.
About twenty years ago, the manager I admired most was Jack Welch, the former CEO of GE. Looking back now, much of what he said may no longer be right. But he was right about one essential thing: the most important thing in a company is its vision. A large company is not managed by rules and regulations. It is held together by vision.
What is vision?
Vision is not a slogan hanging on a wall. Vision is what you do, not what you say. It is how you actually operate. I no longer remember Welch's exact words, but that was roughly the idea.
So how do we manage so many people? How do we organize ourselves? In truth, we do not really have an organization in the conventional sense. We are organized by vision. We use a shared vision to bring people together. That has advantages, and it has disadvantages.
In the future, we will try to keep the strengths and compensate for the weaknesses. But this is our character. We are not managed by KPIs or formal evaluations. We are driven only by vision. That vision is not even written down. It has never been written into any document. It lives in the way we do things and in our attitude toward the world.
Different people in the company may understand the vision differently. Each person's version may not be exactly the same. But in the broad direction, we are aligned.
I think the heart of it is still this: we approach the world with goodwill, and we want to do something meaningful. That is what holds us together.
I will speak first, and then we can take questions. I will probably keep coming back to this vision, because it is real. It is not a story we invented. We really think this way, and we really act this way. Otherwise, many of the things we do would be impossible to explain.
Why are we so committed to open source?
Because the vision itself requires open source. Without that vision, you cannot organize people this way. Some companies also open source, but their open source is different from ours. Their open source has a sense of being forced. It does not feel like what they truly wanted. For us, it is exactly what we wanted. From the very beginning, we were very clear about this.
First, there is vision. Second, we believe that if AI is to succeed commercially, open source is beneficial. That may sound contradictory, even counterintuitive, because historically open source and commercialization have often been in conflict. But AI is different from previous software.
Historically, a software company might have a market worth only a few billion dollars a year. If it open sourced its product, that market might shrink to tens or hundreds of millions. But AI is large enough that, in the end, it may account for something like 10 percent of human GDP. That is an enormous number.
No single person or company can monopolize something that large. You have to share it with others, or you will not survive. This is different from open sourcing a smaller piece of software, because the market was never that large. AI is simply too big. If we try to monopolize all of its benefits, history will discard us. I think this is an objective law, a view of history.
It is not that if I keep everything closed, I can monopolize the market. In theory, that does not match reality. You will inevitably meet resistance. Other forces will find ways to prevent you from achieving that goal. In that situation, you do not necessarily need to think according to traditional business logic. You need mechanisms that ensure the benefit you take for yourself remains limited. Only then might you actually succeed.
You need restraint. I believe restraint is necessary.
If we want to bring AI to fruition in our hands, the first requirement is restraint. We cannot think, "What percentage of human GDP will belong to me?" or "What percentage of China's GDP will belong to me?" The more you think that way, the less likely you are to succeed.
From the beginning, we felt restraint was necessary. The more restrained you are, the more likely you are to accomplish this. That is a business consideration, though at a macro level. It fits my intuition, at least. Or, more precisely, this is what I genuinely believe.
We do not have many special advantages. We are not especially capable. We are not richer than others, and we cannot say our people are better than those at other companies. Not really. When we founded this company two years ago, we did not have a lot of money, many chips, much reputation, or much ability to rally people. We were just a group of ordinary people.
Truly, we were ordinary people. I prefer the narrative that a group of ordinary people did something extraordinary, rather than that a group of geniuses did something extraordinary. That is closely tied to restraint and to our vision.
Will open source conflict with commercialization? In AI, I believe if you lack restraint, you cannot rise. Open source is part of restraint. And our restraint is not limited to open source; it appears in many areas. In general, we do not treat open source or restraint as burdens. The more restrained you are, the easier it may be to succeed. At least so far, this has been borne out.
Otherwise, how could we explain what we have achieved? We had no special weapons. Our starting point was low. Our resources were few. Our people were essentially a random group of ordinary people. I myself was just a college graduate, and not from the very top school. Restraint is part of our vision. AI is too large, and the benefits are too large.
As long as we can make it happen, the eventual benefits will be enormous. Even a small share will be enormous. So right now, there is no need to think about exactly which piece of the profit we should take or how we should take it. The pie is large enough. A tiny slice is already enough.
That is why we have said before that we only want to earn a reasonable profit. What matters is your intention, not the absolute size of the profit.
This is not abstract rhetoric. It is reflected in our API pricing. For us, a reasonable profit means roughly this: we buy a batch of servers, and we recover the cost in about ten months. Given the risks and the upfront investment, even if we depreciate a server financially over three or five years, commercially we think a ten-month payback is enough. That is the logic behind our current API pricing. For V3.2 Flash and other models, the standard is the same: recover the cost of the equipment in ten months.
This is not profit maximization. If we wanted to maximize profit, we would set prices higher. At this price range, user demand is inelastic. If I raised the price by 50 percent, or even doubled it, token consumption would not change much. If I doubled the price, revenue would almost double.
Let me tell you a story. For one of our models, we were initially worried demand would be too high, so we priced it relatively high. People inside the team were not happy. Later, I cut the price to one quarter of the original level, and everyone became happy. That, I think, reveals what we really believe.
The point of all the work we put into making the model good was to make it extremely cheap, extremely capable, and widely usable. When we cut prices, many people in the company group chat cheered. That is our motivation. That is our vision. That is the internal consensus that lets the company cohere around this work.
This is unusual. For our competitors, a price cut is certainly not good news. If you cut prices in half, ARR falls in half. But for us, ten months to recover cost is already commercially satisfying. Externally, we also believe this is a price users are happy to see. The company and society both benefit.
Someone just commented that a ten-month payback still implies high profit. That is true. There is still room to reduce prices, and there is also room to optimize the models further. Overall, the room for price reduction remains large. But the point is that we can achieve ten-month payback while others cannot. Companies like Alibaba or Tencent do not have our optimizations, so their costs would likely be several times higher.
Why do we not cut prices further? Because there is little elasticity. If we lower prices again, demand will not increase much. Everyone can already afford the current price. People are not choosing not to use it because it is too expensive. Lowering the price further would not bring the company more revenue, nor would it create much more value for society.
Still, the principle remains: pricing is not based on maximizing company revenue or profit. It is part of our restraint. In the short term, a higher price might bring more revenue. In the long term, that is not so clear. To me, restraint is a strategy. Sometimes you give up something in exchange for something larger.
Open source is the same. You can call it pressure, or you can call it giving up profit. Internally, that sacrifice makes us happy. Employees feel a sense of achievement, and the company gains cohesion. Society benefits, peers benefit, ordinary people benefit.
In the long run, I believe this restraint increases our probability of achieving AGI. I have no doubt that AGI will have tremendous commercial value. Given that, my priority is not how to take a slightly larger share. My priority is how to increase the probability that we can actually achieve it.
The same restraint has shown up elsewhere. Last Spring Festival, our consumer users suddenly grew sharply. But we did not focus on retaining those users at all costs, monetizing them, or grabbing that commercial opportunity. We did not rush to compete for users or squeeze money out of them. We simply worked hard to serve them well.
We never thought, "We must become the next super app," or "We must compete with so-and-so," or "We must become the next ByteDance or Tencent." We could have tried. We did not.
I see that, too, as part of restraint. Do not try to make money from everything. Once you have users, you might imagine you can become the next ByteDance and eat that whole market. Commercially, that might be feasible. Last year, if we had spent heavily to compete with ByteDance for users, that would have been one path.
But we chose restraint. We chose not to fight for that. There may be watermelons later, and what lies in front of us may only be sesame seeds. We should not grab every sesame seed. Some of them may be large, but compared with what AI may become, they are still small.
Looking back, not pushing hard on the consumer side last year was probably right. The larger opportunities really are still ahead. If we had spent a lot of money last year to make the consumer product huge, what would we have gained? Not much.
These are my real thoughts. The AGI opportunity ahead should be enormous. I do not even need to think now about what position I will occupy in it, or what my business model will be. If the commercial opportunity is truly that large, there will always be a way.
We will still pick up the sesame seeds along the way, but only casually. We will not stop and make them our main business. Last year's consumer DAU was, in my view, a small matter. But we picked it up, and at relatively low cost we maintained user usage, because it may be useful later. Even if we do not know now what those users will be useful for, and even if it is currently a pure cost, it may matter later.
This year, our API and AI-related ARR may also have an opportunity. If demand continues to expand, and if we can buy more GPUs, then reaching several hundred million dollars in ARR is very possible. If AI revenue reaches one billion dollars, our cash flow could basically turn positive and cover R&D and all expenses.
That is possible, but we have not made it our priority. It is something we will do, and it matters. But it is not our first priority. The bigger opportunities are still ahead. Last year's consumer opportunity and this year's B-side opportunity are things we should do and do well, but they are not our goal. Most people in the company do not see them as being remotely as important as AGI.
On Open Source
Let me say more about open source, because many people have asked about it.
First, we will open source. Our strongest models will probably also be open sourced. I do not see the benefit of closed source. I do not see an inevitable benefit.
ByteDance's model is closed source. What advantage does that bring? I do not see it. Even if a model is open sourced, even if you tell everyone everything, the threshold remains extremely high. For others to use it well is hard. For them to use it and also bring the cost down is even harder.
It is not true that once I open source, others can easily achieve the same deployment cost. There is still a lot of work involved. Even if people understand the principles, not every company has the willingness or capability to organize people to reach that goal. Many companies are not good at it. There is too much resistance, and too many managerial and physical constraints.
This is also an advantage of a startup at our scale. If you are too small, you do not have enough force to do this. If you are a large company, it is hard to organize. Our current scale is a sweet spot.
Open source does not affect our business model. The premise is that we only earn what corresponds roughly to a sixfold profit, with ten months to recover cost. Under that model, open source has no impact. If you want to earn one hundred times profit, then yes, open source will affect that, because third parties can deploy the model themselves. Even if their cost is twenty times yours, they can undercut you.
Is this sustainable long term? I think it is. Under our vision, open source is sustainable, and we intend to keep doing it. You can see it as restraint; you can also see it as making a longer-term trade. This strategy gives us more room at the frontier of technology and increases our probability of achieving AGI. It makes us calmer.
Think about it: we basically do not need to work overtime, because it is not that hard. For other companies, it may feel very hard, because they are trying to think about too many things. From the outside, it may look as though we chose the hardest path: research, frontier technology, the hard mode. But in other areas, we gave up a great deal, and that makes us powerful. It lets us move lightly.
So my judgment is that open source is sustainable. There is no conflict between open source and paid commercialization, as long as we are talking about a reasonable multiple rather than extreme profit. A sixfold profit sounds high, but in today's AI efficiency environment, it is not particularly high. In the future it may fall to four times or three times, but it will still be a large profit.
I am not worried about others deploying our model and competing with us. Not at all. In fact, we hope they can deploy it. We will provide as much help as we can to the open-source community, so that people can deploy our models successfully. I do not worry that they will steal this business from us. The market is big enough. What I worry about is that they fail to deploy it, miss details, get worse performance, or incur higher costs.
The open-source model we provide and the model we deploy ourselves are the same. We will not open source a worse model while using a better one internally. Last year, on the consumer side, we were basically open source the whole time, and we did not see conflict. We really did not.
The Long-Term Vision: AGI
Our long-term vision is AGI. Different people may define AI differently, but that does not stop us from treating AGI as our goal. From a technical perspective, the roadmap to AGI is relatively clear.
With the current generation of AI technology, if you can describe a problem clearly and give the model complete context and instructions, it already surpasses humans. But the premise is difficult: complete context and complete instructions.
In today's meeting, for example, we share a long and rich context. Everyone here may carry decades of context. AI does not have that. In a limited context, it can outperform humans, but it still cannot replace a human employee.
If you hire an employee, that person may spend two months learning the company, the environment, and the work. After that, they understand what you mean when you say, "Ask Xiao Wang to come over." They know who Xiao Wang is. An AI does not have those two months of accumulated context. You would have to tell it who Xiao Wang is, what his role is, where he is, how to find him, and what to pay attention to.
So AI can do the task only if you provide all the context, but in reality you cannot provide all the context. That is why AI cannot yet replace an employee.
If AI had the ability to learn continuously, like an employee who spends two months learning the company, then it could replace people across many tasks. The next missing step is learning how to learn.
AI development can be understood as a staircase. Last year's step was chain-of-thought. We found that by letting the model think through problems, intelligence could reach a higher level. That raised the ceiling.
This year's step is the agent. With agents, the range of tasks becomes much broader, and the intelligence ceiling rises again. It is a staircase because each step depends on the previous one. Agents need chain-of-thought. Chain-of-thought needed the previous step, the language model. No step is wasted.
After agents, the next problem we believe must be solved is continuous learning: how to let models keep learning over time, rather than requiring a large, separate training process. This is related to completing tasks; it is part of the same underlying problem.
Standing where we are now, at the agent stage, the next bottleneck is visible. It is continuous learning. We have to cross that obstacle, and there must be a way across, but it will take time.
After continuous learning, we may arrive at something like a singularity. Once a model can continuously learn, it can do all the things humans can do. It can develop its own next version, conduct research, and create better AI models. That is why people call it a singularity.
But I do not think it is truly a singularity. It is likely a gradual process, not a sudden rupture. We use the word out of habit, because early forecasters predicted a singularity there. In reality, it is continuous.
After that step, I think embodied intelligence comes next. Our ideal roadmap is: first solve learning to learn; then reach the self-iteration point; then move into embodied intelligence. Once embodied intelligence arrives, AI enters the physical world. It can do housework, care for the elderly, and meet concrete human needs.
Others may see the roadmap differently. There is no absolute right or wrong. We simply think this route is the easiest, because each step requires the least new work. If you solve continuous learning first, then self-iteration, and only then embodied intelligence, the later steps can be developed with help from the earlier technologies. After the self-iteration point, embodied intelligence may no longer need humans to build it; the model may be able to build it itself.
That is our long-term goal. That is what we mean by AGI.
Commercialization as a Byproduct
Back to reality. Last year, the reality was that everyone wanted to build chatbots and compete for consumer traffic. This year, everyone wants B-side revenue, because if you are not participating, you are not at the table.
But internally, what we really care about is the AGI roadmap and the next technical breakthrough. A strange thing often happens: the thing you most desperately want is often hard to get; the thing you do not obsess over often comes more easily.
Our strategic advantage is that we think about AGI and work on AGI. When we build applications, consumer products, or B-side services, we do not need to spend that much effort. Standing on a higher technical platform, we can do lower-level applications with much less force.
Last year on the consumer side, that was exactly what happened. We did not spend much energy on consumer products. At one point, we almost did not want to maintain those users, but they would not leave. This year, B-side revenue also looks fairly optimistic. I suspect the numbers compare well with peers.
But we did not build this as a separate business. It was incidental. On the road to AGI, serving models through an API is a step we must pass through anyway. Making that API available commercially is a byproduct. We only need a few people to maintain it. There is no customer service team, no sales team. Users come by themselves.
Both consumer users and B-side users are byproducts on the path to AGI. We do not build AGI in order to serve consumers or businesses. We build AGI because we are pursuing AGI, and if the process happens to produce something commercializable, we commercialize it.
That is different from other companies. They build models to serve consumer users or business users. Our original intention is still to pursue AGI.
This creates a kind of dimensional advantage. AGI is a larger vision. It attracts stronger people and creates stronger cohesion. That gives us an organizational advantage.
The advantage is not that our people are smarter. It is about how talent is organized, motivated, and made to collaborate. Putting smart people together does not automatically make them cooperate passionately toward a common goal. You need vision. Our previous experience has taught me that the vision of AGI is powerful.
The Company's Core Interest
We must be restrained in many areas. But what is our core interest?
There is only one: maintaining team stability. That is our greatest core interest. You could even say it is our only core interest.
As long as we can keep the team stable, we will achieve AGI. It is that simple. If people do not leave, and we can keep working, then we can do it. Money is not the problem. Resources are not the problem. Other factors are relatively easy to obtain.
For us, the one thing we cannot compromise on is team stability. It is our biggest challenge and our biggest risk. Recent financing has reduced that risk substantially, because employees received meaningful options. As long as the most important and longest-serving employees remain stable, others are unlikely to leave. Even if others receive fewer options or less income, they may still stay, because they are not here only for money. They want to work in an environment that can achieve AGI.
Historically, our talent turnover has been low compared with peers. But this remains our greatest challenge. Everything else is a matter of time. Other issues may delay us by six months or a year, but they will not prevent us from succeeding.
Much of what we do is designed to preserve team stability. Apart from that, almost everything else can be restrained. We have never wanted to become the enemy of any internet giant or smaller company. I hope we can empower them, help them, and assist everyone in doing this work.
We are willing to help even competitors, including Alibaba, Zhipu, and Moonshot, become better. We lose nothing by doing so. We are open source anyway. If you cannot reproduce something, tell us, and we will tell you how to reproduce it. That is part of open source.
What DeepSeek Will and Will Not Do
Our stance externally is simple: we only work on the main line of AGI. That means the core path of language models, chain-of-thought, agents, and so on.
AI is broad. There are many areas that we do not believe lie on the main line of intelligence, such as 3D or video generation. We probably will not do them. World models, at least at this stage, do not seem closely tied to raising the intelligence ceiling, so we will not do them either.
But if others do these things, we are happy to help. Whether we have the time or people is another matter, but there is no conflict of interest. We hope AI technologies can be used across production environments and industries to improve social productivity.
Our goodwill has not damaged our commercial interest. If anything, it may have helped. That seems counterintuitive, but it is true. If we violated this principle, would we necessarily get more? I do not think so.
At this stage, the most important thing is training AI well. Training AI well does not require a world model, and it does not even require multimodality. If you narrow the training scope, some tasks cannot be done, but the core algorithm still holds. Multimodality will eventually need to be done, and we are doing it, but it is a component rather than the main line.
Video generation became very popular after Sora. Everyone seemed to think that if you did not do video, you were not an AI company. But if you think carefully, it has little to do with the roadmap of intelligence. Commercially, video generation may be a good business. But we will not do something just because it is a good business. We do something because it lies on the roadmap of intelligence.
Resources, Chips, and the Gap With the United States
What do we lack compared with the United States? The main gap is resources. We do not have enough chips.
We currently have roughly the equivalent of 20,000 H-series GPUs, most of which arrived only in the last month or two, and some machines have not yet arrived. Last year we had relatively little compute; this year we are expanding aggressively. Over the next few months, more machines will arrive, mostly Nvidia.
How many chips do we need? As many as possible. Within what we can afford, more chips are better. Our strategy is to buy as many as we can at reasonable prices. If we could spend all the financing within six months, I would consider that ideal. Turning money into Nvidia GPUs is better than leaving it in the bank.
The main worry is not money; it is not being able to buy enough chips.
The gap with the United States is mainly resources. In people, the gap is not large. It is essentially the same population of talent. Some Chinese researchers remain in China; some go abroad. It is not that all the smarter people go abroad. China does not lack talent, and our base is large.
The talent gap that does exist is also fundamentally caused by the compute gap. With less compute, we have fewer opportunities to run experiments, and fewer experiments slow talent development.
At the scale of the largest models, we cannot afford to train them. Even if we spent 50 billion yuan, we still could not really train and use them properly. The largest models today have around 800B activated parameters. In China, we are still at the scale of tens of billions of activated parameters.
To train a model of the same size as the frontier U.S. models, we might need 50,000 GB300s or 200,000 Huawei 950 cards. That is just for training, not research. Our current and near-term resources are enough for us to do more experiments at the tens-of-billions activated scale, but we are still far from training an 800B model.
So the gap with the United States is a resource gap. Many other gaps - talent, model capability, applications - can be understood as consequences of the compute gap.
Where Model Competition Will Ultimately Differ
In the long run, the differences between large models may not be very large. The final differences will likely appear in three areas: cost, time, and user experience.
Cost is easy to understand. If two companies provide the same service at the same quality, the question becomes: at what cost can they provide it? Cost will be a real barrier, and probably the most important difference.
The second is time. Whether you can achieve something a few months earlier or later matters.
The third is user experience. There will be some differences in stickiness and experience, though perhaps not the most fundamental ones. The fundamentals are cost first and time second.
For now, the best thing to spend energy on is still AGI: pushing intelligence forward and raising the lower bound of capability. That produces a higher return than expanding product lines or designing more commercial paths. In the past three years, at any given moment, talking about product lines and commercialization paths would have been mostly a waste of time, because things changed too quickly.
We have always been commercializing. We have consumer users and B-side revenue. But we have not taken commercialization as the goal. The point at which we would fully turn toward commercialization is still far away. The biggest returns still come from extending the technology and building the next generation.
I hope many commercial opportunities can be pursued by other people. How to use AI should be something society and partners do together, sharing the benefits, rather than something we try to swallow alone. We do not have the energy or the organizational scale to do everything.
Decision-Making and Organization
Our company is built on consensus. I do not decide everything alone. I seek consensus. My authority and influence inside the company are based on consensus.
If I want to do something, I first look at what everyone thinks. Do people want to do it? I may guide or lean in a direction, but the effect of guidance is limited. It still has to rest on consensus.
Our management has two lines: top-down and bottom-up.
Bottom-up means everyone decides what they want to explore. No one controls them. There are no KPIs.
Top-down is what we call "serious work": when the company collectively needs to do something, such as release V4, people divide responsibilities and work together. Ideally, serious work should not take more than half of an employee's time. The other half should remain unassigned, so people can explore what they believe matters.
We generally do not work much overtime. One reason is that research requires a relaxed environment. If people are pushed too hard, they cannot do research well. The second reason is focus. Because we are extremely focused, there are not that many things to do. Restraint means we simply choose not to do many things. You can see that some of our products are incomplete, and we have not rushed to fill every gap. That is also part of our culture.
As the company grows, the organization will need to change. Some departments will need more rigorous hierarchy. Others can remain looser and flatter. We are already making those adjustments, because without them some work cannot move forward.
We do not have a model to imitate. Every step has come from our own real situation. We made decisions based on reality and on what we believed was optimal. We are a product of this era and of our circumstances, not the result of copying anyone.
We are not Bell Labs. Bell Labs did not need to commercialize in the same way. We do. We have to survive. We may have a very ambitious mission, but fundamentally we are still a company. The B-side business matters because in the future we may need it to survive.
Many great companies have had pursuits beyond profit. Those pursuits did not prevent commercialization; in fact, they sometimes made commercialization better. We are still a company. We simply make choices about which money to earn, when to earn it, how much to earn, and what to rely on.
China's Role and Domestic Chips
In the global division of labor around AI, Chinese companies may ultimately play the role of producing at the largest scale. We have large manufacturing capacity, chip capacity may eventually be large, and we have abundant electricity. In the end, China may become one of the major powers in AI.
Chinese companies will likely make AI products cheaper. In many industries today, Chinese products are not dramatically different in quality from American products, but they are cheaper. AI may follow the same pattern: Chinese AI may be systematically cheaper, just as Chinese services in other industries are often cheaper.
Domestic AI chips now have a historical opportunity. In the past, the difficulty with domestic chips was ecosystem. You could buy the card, but you could not use it easily. Nvidia's CUDA ecosystem was a strong moat.
That is changing. The CUDA moat is being dismantled quickly for three reasons.
First, AI makes it much easier to build an ecosystem, because AI can write code. Second, there are new technologies. We have a technology called TileLang, a high-level language. With TileLang, CUDA operators can be written quickly, and the whole ecosystem can be rewritten with much less code. Combined with AI, the path looks unobstructed.
Third, CUDA evolved from gaming cards. Historically, AI compute was a smaller market than gaming, so that coupling made sense. Now compute cards are larger than gaming cards, so there is no reason the two must remain coupled. Future dedicated AI chips, whether from Huawei or Nvidia, will not be bound to the old CUDA assumptions.
We believe that within a year, one thing will be validated: the domestic chip ecosystem is not a problem. The perception that domestic chips are unusable or hard to use will be reversed by facts. The problem will be capacity, not ecosystem.
We mainly cooperate with Huawei. Huawei does its own adaptation, and we will deeply participate in the ecosystem. Huawei's issue is still capacity. The 16,000 Huawei cards allocated to us are equivalent to only about 4,000 Nvidia B-series cards. That is not enough to train a next-generation model, but it is useful for helping Huawei improve the ecosystem.
Our main work on Huawei adaptation is to make the high-level language compiler good and to make TileLang good. Once TileLang is good, many problems will resolve naturally.
In V3 training, we used Nvidia cards but no longer relied on Nvidia's ecosystem. We built a high-level compiler, TileLang, and built the rest on that ecosystem. If we redo that process on Huawei cards, then the adaptation is complete.
Huawei's 950 supernode can, in performance and price, substitute for Nvidia's GB200 or GB300. The price will be higher, but within limits. Even if it is 100 percent more expensive, that can still be considered a viable substitute. The main tradeoff is that four Huawei cards equal one Nvidia card, and Huawei is about two years behind. So the chip gap is roughly four times plus two years.
We hope we do not need to vertically integrate upstream into chip design. If you operate a power plant, you do not necessarily need to manufacture generators. If the equipment can be bought at a reasonable price, why make it yourself? AI is large enough that we only need to do one piece: the piece we are best at and believe is most central.
Scaling, Cost, and Multimodality
We believe in scaling. The larger the scale, the better the results, and the more capabilities can be unlocked. What prevents us from scaling is compute, not lack of desire.
We have not touched the scaling wall. The models we train are this size not because we believe this size is enough, but because this is what our resources allow. Silicon Valley may talk about scaling reaching its limit, but for China we are still far from that point.
Multimodality is important for products and for consumer user experience. We are doing it, and future V4-related versions should support native multimodality. But for intelligence, multimodality is a component, not the main line. Search is also a component.
The next generation model, in my view, must have continuous learning. Before that, what we can do is improve cost, quality, and speed. But a major breakthrough requires continuous learning.
Why do we care so much about computational efficiency? Not every company does. A purely commercial company may not have a strong incentive to make models highly efficient. Lower cost can reduce what you are able to charge.
For us, low cost is part of the vision. Our colleagues are ordinary people; they know users have to pay. They empathize with users. If the service is cheaper, more people can accept it. That is why many of us want to keep lowering cost.
There is also a technical reason. The lower the cost, the larger the model we can afford to train. Under limited compute, higher efficiency lets us support larger models. Large companies may solve problems by simply adding resources. We prioritize cost efficiency.
Data, Post-Training, and Hallucination
Data is extremely broad. In some sense, data is half the model.
At this stage, solving AI depends heavily on data labeling. You could say that half of our core researchers are labeling data. The cost of high-quality data labeling in China is not meaningfully lower than in the United States, especially for high-end data. This makes it difficult for us to invest in the same way U.S. companies do.
We take a two-legged approach. Some data is cheaper to label, some more expensive. We start with the lower-cost data. High-quality data is not just a matter of money. The bottleneck is time. OpenAI and Anthropic started earlier, had more capital, and had more chips. In China, you could say serious work in this direction began only in the past half year. It will take time, but within a year I think domestic high-quality data should improve significantly.
Hallucination can be improved through better post-training. It is solvable, or at least improvable. People have not necessarily put enough effort into it. For us, hallucination is a problem, but we may categorize it more as a product problem. We will solve it, but it is not the central problem.
Continuous Learning
Investors may see agents as the main topic right now. Researchers, however, are looking more at learning: how to solve the learning problem. Learning may not be a single technology. It is a problem, and many technologies may contribute to solving it.
The difficulty is that no one has yet found a method that clearly works. The whole world is exploring. We have many promising ideas, but none has fully broken through.
Inside DeepSeek, one narrative is especially important: the next model should help us develop the next model. The first goal of our model is not that everyone else finds it useful. The first goal is that we ourselves find it useful. If it improves DeepSeek's own efficiency, then we can develop the following model faster.
This may sound strange, but many of us think this way. The model should first help us achieve AGI. We now clearly need AI to help us achieve AGI. Even if it is not yet autonomous and still works together with humans, it is already very useful.
If continuous learning is solved first, general intelligence becomes much easier. An AI that can continuously learn will greatly improve our own research efficiency. Without that, building general intelligence manually is exhausting, labor-intensive, data-intensive, and not very cost-effective.
AGI will not have a single sharp threshold, but the process will be nonlinear. We believe AI can accelerate AI research. That means progress may become nonlinear later.
Domestic Hardware and Timing
On the current AI paradigm, China may reach performance comparable to foreign models within one or two years, perhaps even this year. But that is not yet AGI. At minimum, AGI needs continuous learning.
Domestic hardware may need several years. First, China must solve the ecosystem problem, which is partly a confidence problem. Then it must solve capacity. I do not believe that five years from now we will still be stuck on capacity. This year, next year, and the year after, capacity may still be a constraint. But five years out, I am more optimistic.
Nvidia cards can basically be depreciated over five years. Huawei cards should probably be depreciated over at most three years, because they already lag Nvidia by about two years and will become power-inefficient sooner. Still, if B200s can be purchased at a reasonable price, they are worth buying.
Our compute lag can be offset in three ways. First, we accept that our model may be smaller and somewhat behind. Second, being behind gives us more time to use clever methods. Third, we aim to rewrite the narrative: instead of being one or two years behind using one twentieth of the compute, we want to use a fraction of the compute and shrink the time gap to six months or three months. In some areas, we may even surpass them. But with an order-of-magnitude gap in total compute, full-spectrum superiority is unrealistic.
Applications and Vertical Agents
In China, it is still hard to judge what the eventual commercial model will be. At this stage, the most reasonable path is to focus on a general agent. Other agents, including finance or medical agents, should have lower priority.
Coding agents should come first, because they can do many things and support many vertical agents. Right now, the most important application is still the coding agent.
On model capability, at the 50B activated scale, I expect our open models to end up not too different in speed and quality from the current wave of open-source models. But there will still be a large gap compared with much larger unpublished models. Closing that gap will likely require a bigger model, perhaps around 150B activated parameters. Optimistically, we may begin training such a model by the end of this year, or otherwise next year.
On Taste and AI Self-Improvement
If AI can self-improve, will taste and intuition still matter?
AI does not lack taste or intuition today. If you ask it to write an essay, its taste and intuition are not the issue. What it lacks is continuous learning.
For embodied intelligence, yes, we will eventually have to enter that stage. For ordinary people, the need is not a computer. People need food, clothing, housing, transportation, care, and everyday physical help. If the goal is to liberate human labor, embodied intelligence cannot be avoided.
Before embodiment, our definition of AGI is practical: can it help us iterate the next model? After embodiment, the goal is similar: can it help us iterate the next generation of embodied systems and robots?
TileLang and the End of CUDA Dependence
A question was raised about whether moving from CUDA and PTX toward TileLang would reduce inference efficiency in the short term.
It does the opposite. It significantly improves efficiency. This is an opportunity.
Previously, you could not escape the CUDA ecosystem. Now we can abandon that ecosystem and use a simpler method, TileLang. It is a high-level language, the programs are faster to write, and the amount of code is much smaller. We can rewrite the whole stack.
This is a major opportunity created by technological development. We are also using AI to write TileLang, though right now TileLang is still written by humans. Even so, it is already much faster than writing CUDA directly.
At the hardware execution level, the efficiency loss is around 1 to 2 percent, which I consider acceptable.
Closing
DeepSeek's path is not modeled after another organization. It is not simply Bell Labs, nor is it a conventional commercial company. It is a company pursuing AGI, with commercialization as something necessary but restrained.
Its central logic is unusually simple: keep the team stable, stay focused on the main line of intelligence, keep costs low, share enough with the ecosystem, and resist the urge to grab every nearby opportunity.
In Liang Wenfeng's framing, restraint is not weakness. It is the strategy that makes the larger ambition possible.
