4 hours, 118 responses, Liang Wenfeng addresses everything in internal communication

Compiled by｜Gu Lingyu, Tencent Technology

DeepSeek recently completed its first round of external financing since its establishment. The total amount raised in this round exceeded 50 billion yuan (approximately 7.4 billion USD), with a pre-investment valuation of about 367.5 billion yuan (approximately 54.3 billion USD). Among the investors, DeepSeek's founder Liang Wenfeng personally contributed 20 billion yuan, Tencent invested 10 billion yuan, CATL invested 5 billion yuan, and NetEase, JD.com, and IDG Capital each invested 3 billion yuan, while the National Artificial Intelligence Industry Investment Fund contributed 1 billion yuan.

Prior to this, Liang Wenfeng had proposed the principle of "no financing, no listing, no commercialization." This large-scale financing marks DeepSeek's official entry into the capital market and has sparked widespread attention in the industry regarding its commercialization path and technological vision.

At a recent investor communication meeting, Liang Wenfeng elaborated on DeepSeek's organizational culture, open-source logic, technology roadmap, and views on the competitive landscape of the industry.

The following is a compilation of nearly 4 hours of Liang Wenfeng's speech during the communication meeting obtained by Tencent Technology, categorized by theme, totaling 118 entries, with the text preserved as much as possible in its original meaning, with only slight editing.

01 Vision and Restraint

1. When we first started this company, we did not think about how much money we would ultimately make, whether we would go to the capital market, or how we would list. The initial few dozen people never thought this way; if they had, they wouldn't have joined.

2. We are doing this with a great goodwill towards the world; we believe this is useful for humanity, and it is something beyond money. Our original intention, our vision, and the vision we have maintained until now are not based on maximizing commercial interests.

3. Managing a large company relies not on your rules and regulations, but on vision. Vision is not a slogan hanging on the wall; it is about how you do things, not how you say it, but how you actually operate.

4. We are unorganized; we are driven by vision, organized around a vision. We do not operate with a mindset of "I need to achieve certain KPIs, and there are no assessments," but solely based on vision.

5. This vision is not even written down; it has never been documented. This vision is reflected in our methods of doing things and our attitude towards the world.

6. We do not have many other advantages; we do not have any special abilities, we are not richer than others, and we do not have better personnel than other companies. In fact, we are just a group of very ordinary people. When we established this company two years ago, we had little money, few cards, no fame, and no influence.

7. The more restrained you are, the easier it may be to succeed, or at least so far this has been validated and can be explained. Otherwise, there is no way to explain why we have been able to succeed: we have no weapons, a very low starting point, and very few resources; our people are just a random group of ordinary individuals.

8. AI is too big, and the benefits are too great. We are very restrained; as long as we can succeed, the benefits will be very large in the end. If you share a little, the benefits are already substantial, so there is no need to consider which part of the benefits to take or how to take them; I think there is no need to consider this matter because the benefits are already large enough.

9. Last year during the Spring Festival, we suddenly had many users, but we did not pursue retaining these users, or monetizing them, or competing for commercial interests. We did not compete for users or make money, but we worked hard to find ways to serve the users well.

10. We do not have the idea of becoming the next super app, nor do we want to compete with anyone, nor do we want to become the next ByteDance or Tencent; we have no such thoughts at all. I believe the upcoming AGI opportunities should be very large; the opportunities for AGI will always be very large.

11. Restraint is a strategy. It is about sometimes sacrificing some things to gain more of others. The decision not to open-source is similar; it can be seen as our pressure or as our concession.

12. This kind of restraint, I understand, can increase our probability of achieving AGI in the long run. When considering a matter, I have no doubt that AGI will have significant commercial value. Therefore, based on this, my priority is not how to increase my share or take more; my priority is how to increase the probability of achieving it.

13. We have always been very restrained and unwilling to become competitors with any major or minor internet companies. I hope I can empower them or assist everyone in doing this, hoping to help everyone achieve this.

14. I believe that by adhering to this attitude, we have not missed out on anything; we have not received less because we open-source or because of our goodwill or our help to others. On the contrary, it may have added points. This seems counterintuitive, but it is indeed the case.

15. We aim for AGI, but we have always been commercializing, which is why we have C-end users and B-end revenue. From historical experience, this strategy has been successful.

02 AGI Roadmap

16. If you can describe a problem clearly, providing complete context and instructions, it has already surpassed humans. But there is a definition, a premise: you must provide complete context and complete instructions.

17. AI cannot replace your employees. But if AI has the ability to learn continuously, it can replace everyone after learning for two months in the company, so we are still a step away from that continuous learning.

18. The development of AI can be understood as a staircase. Last year's step was the chain of thought. We found that using the chain of thought can elevate intelligence to a higher level.

19. This year's step is the Agent, because we found that using Agents can accomplish more tasks, expanding its capability range and raising its intelligence ceiling. Agents need to use CoT, and CoT also needs to use the previous steps, which are language models, so no step is wasted.

20. After Agents, we believe the next problem to solve is continuous learning, which is how to enable models to learn continuously, rather than requiring strong training; it should be able to learn continuously over a long period like humans.

21. After continuous learning, we may reach a singularity. This singularity is when the model can continuously learn and can do everything humans can do. It can develop its own versions, conduct its own research, and develop its next version, creating more advanced AI models.

22. This singularity is not a singular point; it is also a gradual process. This process may also be a relatively long gradual change; it is not a sudden mutation. But habitually, we all think it may be a singular point.

23. This is our speculation; we believe the timeline should be: first solve learning, then reach that intelligent singularity, capable of self-iteration, and then achieve embodied intelligence. Once we have embodied intelligence, it can enter the real world, do housework for you, and take care of the elderly.

24. If we first solve continuous learning, then the self-iterative singularity, and then embodied intelligence, the journey will be much easier. Because later on, you can use earlier technologies to help develop later technologies.

25. We only focus on the main line of AGI. The AI field is vast, and there are many things we believe are not on this main line, such as 3D and video generation; we think they may not be closely related to the main line of intelligence, so we will not pursue them.

26. Video generation was very popular when it first came out; it seemed like a must-do, and if you didn't do it, you weren't an AI company. So I find it strange; if you think about it carefully, it has no relation to the roadmap of intelligence.

27. In business, it is a good business, but it has nothing to do with intelligence. We will not pursue it just because it is a good business; we will only pursue it if it is something on the intelligence roadmap.

28. From our judgment, world models and intelligence are not the most important things at this stage. The most important things are AI training and how to solve continuous learning after AI training. This is our company's judgment, of course, every company's judgment is different.

29. We currently believe in a narrative that AI can accelerate AI research. That is to say, it is not linear, because you can use AI to accelerate your own research, so it may become non-linear later.

30. I believe embodied intelligence must enter; ultimately, it must be embodied. Because for a normal person, their needs are not for a computer, right? Because normal people need food, drink, entertainment, clothing, and shelter; they do not need a computer. They need embodied intelligence to solve specific human needs.

31. What do we hope AGI can do? It can help me iterate the next version of the model, just like that. If we have embodiment, we hope it can also iterate the next version of embodiment, to create the next version of robots.

32. The core capability of the next generation model must include the ability for continuous learning; only then can it be called the next generation model. Before that, what we can do is reduce costs, improve effectiveness, and increase speed. But for a major breakthrough, it should possess continuous learning.

33. The current limitations of Agent capabilities are due to its inability to learn continuously and effectively. If we can first complete continuous learning, then AI's capabilities will be very strong, significantly enhancing our research efficiency.

34. If continuous learning is achieved, general intelligence may become very easy to achieve; using it will be very straightforward. So I say this is a result we hope to see; it will save us effort and make things easier. Otherwise, if you want to manually create general intelligence now, it is a tiring and labor-intensive task, data-intensive and human-intensive, with low cost-effectiveness.

03 Team and Talent

35. Our previous experiences have taught me that the vision of AGI is very powerful. This talent advantage is not about having smarter people than others, but about how to organize, motivate, and collaborate with these talents.

36. Bringing smart people together does not mean they can naturally collaborate and passionately pursue a goal, so you need a vision.

37. Our greatest core interest is to maintain team stability. This is our greatest core interest, and it can even be considered the only core interest. As long as I can maintain team stability, I will definitely succeed; it is that simple.

38. Money is definitely not an issue, resources are not an issue, and other factors are easily obtainable. For us, there is only one core interest, one that cannot be compromised: we must maintain team stability.

39. This is also a significant challenge we face, or rather, I think it is the biggest risk. Of course, this risk has been significantly alleviated with our recent financing, as everyone has received a considerable amount of options.

40. From the perspective of team stability, as long as the most important and longest-serving employees can remain stable, others are unlikely to leave. Even if others have slightly fewer options or lower income, they will not leave. Because they are not all driven by money; everyone hopes to work in an environment where AGI can be achieved.

41. Everything else is just a matter of time; the most it can cause is a delay of half a year or a year, but it will not mean we cannot achieve it. We definitely do not lack money, and we definitely do not lack resources; in fact, we do not lack these.

42. The gap between us and the United States mainly lies in resources; the gap in personnel is not very significant. There is almost no gap in personnel because it is the same group of people, possibly Chinese. When Chinese people go abroad, some stay in the country, some go abroad, and it is not that only smart people go abroad.

43. Talent is not the bottleneck; resources are the biggest bottleneck. Resources primarily affect talent cultivation because of limited computing power, we have fewer experimental opportunities, so our talent overall lags behind the U.S. The talent gap is essentially due to the gap in computing power.

44. The shortage of AI talent is also temporary, and we have already seen it significantly alleviated. Because there is no shortage of AI people; every company will quickly cultivate talent, and training people is fast.

45. There are currently too many companies in China working on models; it is still too many. The U.S. may have three companies, while China has too many companies working on foundational models. Ultimately, there will not be a need for so many people to work on foundational models; it will definitely converge.

46. Our company's management actually operates on two lines: one is top-down, and the other is bottom-up. The bottom-up approach allows everyone to do what they want without anyone managing them or having KPIs.

47. Generally, we hope that employees have half of their time unassigned, allowing them to do whatever they want. This is a research scope that allows them to explore based on what they find important, without preconditions.

48. We generally do not work overtime. There are two reasons for overtime. The first is that research requires a relatively relaxed environment. If you push too hard, you cannot conduct research. Since it requires you to have interest, you need to think about these issues in a relaxed environment to explore.

49. The second reason is that we are very focused. Being very focused means we have very few tasks to accomplish. Therefore, I do not have that much to do, and I do not need to work overtime. This is consistent with the previous restraint.

50. Our company is fundamentally built on consensus; I am not the one who decides everything, but I seek consensus. My authority and influence within the company are based on consensus.

51. This decision-making mechanism is essentially a consensus-seeking mechanism; it is not that I can push something forward; it must be consensus for me to push it forward, and then I will push.

52. As the number of personnel increases, we will make adjustments. We should make this adjustment soon because I am already making this adjustment. If we do not make this adjustment, many things cannot be advanced. There are indeed many departments that should have organizational structures.

04 Computing Power and Resources

53. How many cards do we need? The more, the better. Within our capacity, the more cards, the better; this is without a doubt. So our current strategy is to buy as many cards as possible at reasonable prices.

54. In fact, spending so much money is very difficult; it is hard to buy that many cards, and the prices are high. We cannot spend excessively high prices; we must ensure that the prices are reasonable. If we can spend 20 billion this year, then our procurement department will have performed exceptionally well.

55. The biggest gap between us and the U.S. is in resources. On the one hand, computing power resources are hard to obtain domestically; on the other hand, our capital investment is less than that of the U.S. We have significantly less capital investment, and the proportion of talent salaries in this is very low. You see, they offer salaries of 100 million USD, but when calculated, talent salaries still account for a small proportion; the bulk is still computing power.

56. All the differences we see, including talent differences, model capability differences, and application differences, can be attributed to differences in computing power resources.

57. We may lag behind the U.S. by about 12 months, possibly 12 to 18 months, or 6 to 12 months. Simply put, we are lagging behind the U.S. by two years, while using only one-twentieth of their computing power to achieve this.

58. This narrative is that we are one to two years behind but using only one-twentieth of their computing power. In the future, we want to rewrite this narrative, which is that we use only a fraction of their computing power but shorten the time to 6 months or 3 months; I think this is a goal.

59. Scaling, we believe in Scaling; the larger the scale, the better the effect, unlocking more functions. What prevents us from Scaling is actually computing power; it is not that we do not want to Scale, but that we do not have enough computing power to do this Scaling.

60. We train such large models not because I think such large models are sufficient, but because I happen to have that much resource. I calculate based on my resources; the model I can accept and train is determined by that, not that this model is sufficient.

61. When Silicon Valley talks about reaching the limits of Scaling, that is for Silicon Valley; for Chinese people, we are still far from that; we have not reached that level of Scaling. This Scaling includes data scaling, model scaling, and training costs.

05 Domestic Chips and Ecosystem

62. The moat of NVIDIA's CUDA is rapidly being dismantled. On the one hand, with AI now available, establishing this ecosystem is much easier than before because AI can write code.

63. The market for computing cards is now larger than that for gaming cards, so there is no reason for these two to still be coupled. The trend is that in the future, they will no longer be coupled. Therefore, dedicated chips, whether from Huawei or NVIDIA, will be dedicated chips, not the previous ones.

64. There is a historic opportunity for domestic AI chips to replace current ones. We believe that within the next year, we will see something validated: the ecosystem of domestic chips is completely fine. Previously, it was thought there were issues, that they could not be used or were not good, but I think within a year, we can reverse this perception or change it with facts.

65. The hardware and ecosystem of domestic AI chips are fine; the only problem is insufficient production capacity. There are no obstacles for domestic cards to adapt; NVIDIA cannot block it. If I can buy NVIDIA cards in a normal business environment, then domestic replacement is relatively difficult; but in the case where NVIDIA cards are unavailable, everyone will have to turn to domestic chips.

66. During the training of V3, it still used NVIDIA cards, but it no longer relied on NVIDIA's ecosystem. V3 used NVIDIA cards but did not use NVIDIA's ecosystem; instead, we first wrote a high-level compiler called TileLang, and then completed everything else based on the TileLang ecosystem, which has almost completely eliminated reliance on NVIDIA's ecosystem.

67. I am relatively optimistic about domestic computing power. I believe that in this regard, NVIDIA is digging its own grave. Huawei's super node, Huawei's 950 super node, can completely match the performance and price of NVIDIA's GB200 and GB300.

68. Four Huawei cards can match one NVIDIA card.

69. I believe that the gap between us and the U.S. in chips will no longer exist in terms of ecosystem, but in chips, it is four times plus two years.

70. We are currently mainly cooperating with Huawei. Huawei adapts itself, but we will participate in this ecosystem and engage deeply with Huawei. Huawei's issue remains insufficient production capacity.

71. I do not believe that five years from now, we will still be stuck on production capacity issues. We are definitely stuck on production capacity issues now, this year, next year, and the year after; I think we may still be stuck on production capacity issues, but five years from now, I think it may not be the case; I am relatively optimistic.

06 Competitive Landscape and Industry Judgment

72. The final effect of each model will differ based on a comprehensive assessment. Comparing model performance must be done at the same cost; this is meaningful. Because when you compare two cars, you compare cars in the same price range.

73. Is Anthropic currently surpassing OpenAI a long-term situation? I do not think this is long-term; it is definitely temporary. OpenAI and Google will likely continue to alternate in the future.

74. When it comes to the global AI division of labor, Chinese companies are likely to play the role of the largest producer. Logically, we have the largest production capacity, including chips; we may have the largest production capacity and the most electricity.

75. Chinese people will make these products the cheapest, and then in terms of effectiveness, after all, there is not much difference between domestically produced and U.S.-produced goods. In the future, AI may be the same, but domestically produced AI may be cheaper. This cheapness may be systemic, just like other industries where services provided by China are cheaper.

76. The final gap should be in three aspects: one is cost, one is time, and one is user experience. Other than that, there may not be much difference.

77. Cost is definitely a difference; I think cost is the primary distinction. The second is time; when can you achieve it? A few months earlier or later makes a difference.

78. OpenAI initially thought it could monopolize the world, but in reality, it will face many challengers. It will encounter challenges and will not have it easy. The U.S. will face challenges, and in the future, it may also face challenges from China because Chinese people are willing to take less to provide this service.

79. Those who take more will be defeated by those who take less. You do not even need to take a lot; if the vision is to take more, you will be defeated by those whose vision is to take less. In fact, everyone has not made money; it is just a vision. If your vision is to take more, you will lose first and face greater difficulties.

80. For us, we are not looking to take the most profit or maximize returns in pricing, but to earn a reasonable profit. This is an explanation. I believe in this; I am not looking for reasons for this because there is no need to find reasons.

81. I think in many experiential aspects, we may be able to do better than the U.S. In terms of products, our product capabilities may not necessarily be worse than those in the U.S. Costs should also be lower than in the U.S., so China will still have competitiveness.

82. The cost is easy to understand because they do not need to do it, so they do not develop this capability. They definitely do not value this matter as much as we do. We can treat it as a very important matter, but for them, it is not important.

83. Large models may not require two large companies or two small companies; it may already be sufficient. The gap is only two things: one is time, and the other is cost. Therefore, it is not that any company has exorbitant profits; I think there are no exorbitant profits. Those who control costs well earn a bit more, while those who do poorly earn a bit less; that is all.

07 Model Development and Technology

84. About half of the people in our company usually think OpenAI is better. In fact, Anthropic has a first-mover advantage, but this advantage should quickly disappear; it is not a long-term advantage. All three of these companies are very strong, and among them, the efficiency is the highest; the costs they incur and the money they burn should be the least.

85. We have been working on multi-modal layouts. It is very important for products; it is very important for C-end user products. But for the upper limit of intelligence, it is a component; it is not the main line itself.

86. We should support related models; our V4 and subsequent versions will support native multi-modal capabilities. However, for us, multi-modal is a component; we do not consider it intelligence itself.

87. I can only say that I currently do not see an upper limit for the scaling of language models. Our current intelligence level, or that of the U.S., does not show any upper limit.

88. Many people internally think this way: first, it must be useful to ourselves; it must first serve our own needs. Then this is the fastest way to achieve AGI. When it is useful to us, it means it may also be useful to others, but we must first ensure it is useful to ourselves.

89. The primary goal of the models we develop is not for everyone to use well, but for us to use well. First, it must be useful to ourselves. After it is useful to us, I can develop the next version of the model faster.

90. We call this "摸奖" (lottery). The threshold is very low; anyone can try, but what they can achieve may depend on talent or other factors. So there is no need for us to allocate resources. It is just that what sets us apart from other companies is that we spend time discussing this issue, thinking about it, and treating it as an important matter.

08 Commercialization and Pricing

91. Our API pricing is set to achieve a reasonable profit, roughly allowing us to recover costs within ten months after purchasing a batch of equipment; I think this is a reasonable profit.

92. If profit maximization were the goal, prices should be set higher. Because in this price range, user demand is inelastic; even if I double the price or halve it, the difference in token consumption is not significant.

93. For one of our models, we initially worried about excessive demand, so we set the price relatively high at first, which did not please the team. Later, I lowered the price to a quarter, and everyone was very happy.

94. The upper limit for B2B business should still be demand; under the current generation of AGI and AI technology, B2B demand should be limited. It will grow rapidly, but it is not an infinite matter; ultimately, it is constrained by demand, not computing power.

95. I now believe it should be achievable; we should aim for it. If I can have hundreds of millions of USD in B-end revenue this year, along with C-end users, that already establishes a certain commercial foundation. If we have B-end revenue next year and this demand can increase, the company will be close to net profit, possibly already achieving net profit.

96. In the worst-case scenario, selling APIs could support a publicly listed company. If there are no new advancements in technology, and our technology freezes here, then we will fully focus on selling APIs and providing these services well; I think that would be sufficient.

97. Given the current situation, I think the most reasonable approach should be to fully develop general Agents; the priority of other Agents should be lower, including financial and medical Agents. We should first focus on Coding, as Coding Agents can accomplish a lot and there are many vertical Agents. At this stage, we believe the most important is still the Coding Agent.

98. I think low cost is primarily a result. Our models have indeed been moving towards a lower-cost direction in terms of model architecture, which is related to our vision. We still have many algorithmic methods that can further reduce costs.

99. Another reason for reducing costs is that the lower the cost, the larger the models I can train, and the more I can afford larger models. Under the same computing power, if my computational efficiency is higher, I can handle larger models.

09 Open Source Strategy

100. I believe we will open-source, and our strongest models may also be open-sourced. Because I see no benefits to keeping them closed; I see no inherent advantages. Byte's models are closed-source; what benefits does that bring? I see no benefits.

101. Even if the model is open-sourced, telling others everything, the threshold is still very high. For others to use it, the threshold is also very high. It is difficult for them to use; secondly, for them to use it, keeping costs low is also very difficult; it is not that easy.

102. Open-sourcing will not affect revenue. I believe open-sourcing has no impact on our business model.

103. I am not worried about others deploying our models and competing with us; not at all. We even hope they can deploy them. We will provide as much assistance as possible to the open-source community to help everyone deploy our models.

104. When dealing with the outside world, our attitude is: we only focus on the main line of AGI. When interacting with the outside world, we are very willing to assist and help anyone, even our competitors, including Alibaba, Zhiyu, and Moonlight, to do better. Because we do not lose anything; we are originally open-source.

105. Is the open-source model we provide the same as the model we deploy ourselves? Yes, it is the same. We will not open-source a slightly inferior model and then use a better model when we deploy; it will be the same.

10 Data and Post-Training

106. Data should almost equal half of the model. There is also the issue of labeled data. Our data labeling is related to our capital investment. Given our capital investment structure, we cannot afford the costs of so much high-quality data labeling because it is very expensive.

107. The cost of data labeling in the U.S. is not much different from that in China. There is no cost advantage for China in labeling data, especially for high-end data, which makes it difficult for us to invest in labeling data like the U.S. This path is very difficult in China because labeling data is simply too expensive, whether we outsource or do it ourselves.

108. Currently, we are basically walking on two legs. It is not that we cannot label at all, but because some data labeling has low costs and some have high costs. We first label the low-cost data.

109. You could also say that half of the people in our company are labeling data. Half of the core researchers, the most important people, are labeling data. We are focusing on data labeling. Solving the AI problem at this stage relies on data labeling.

110. The bottleneck for high-quality data labeling, I believe, is time; it requires time. Because for OpenAI, for foreign companies, and for Anthropic, they are earlier, have more capital, and have more cards.

111. The hallucination problem of large models significantly affects user experience. The hallucination problem can also be solved with a method, but it is a long-term proposition. The hallucination problem can be considered something that can be improved through better post-training.

11 Organization and Company Positioning

112. First of all, we do not have an object to imitate. Every step is based on practical situations, seeking truth from facts, making decisions based on actual conditions, and finding out what we should do. So it is a product of the times, or a reflection of the reality; it is not a result of imitation.

113. We are clear that we need to commercialize. Ultimately, we still need to survive; after all, we are a company, and the government will not give us a penny.

114. We are essentially still a company; it is just that we consider which money to earn, when to earn it, how much to earn, and how to earn it; we have trade-offs. Many companies achieve greatness because they have pursuits beyond profit. That pursuit ultimately not only does not affect their commercialization but can also enhance it.

115. For partners, this financing is carefully selected. First, I think the interests are quite aligned; they are the ones whose interests align most closely with ours, who have the least hostility towards us, or who most hope we can succeed. Not everyone hopes we can succeed because we still harm the interests of many others.

116. AI currently lacks the ability for continuous learning; it does not lack taste and intuition. AI's taste and intuition are not an issue. If you ask it to write an article, I believe its taste and intuition are fine.

117. We hope to focus on just one area. I think AI is very large and does not require me to… I only focus on one area. If focused, and I believe the business interests here are already large enough, if the AI era produces many trillion-dollar companies, I believe we will be one of them.

118. We hope to support more people, but we do not have that much energy. We have the willingness and there will be no conflict of interest, but whether we do it is another matter. But at least there is no conflict of interest here; we hope for win-win cooperation.