4 hours, 11 topics, 118 responses, Liang Wenfeng answered all questions in the internal communication

DeepSeek has recently completed its first round of external financing since its establishment. The total funds raised in this round exceed 500 billion RMB (approximately 74 billion USD), with a pre-money valuation of around 3.675 trillion RMB (about 543 billion USD). Among the investor lineup, DeepSeek founder Liang Wenfeng personally contributed 200 billion RMB, Tencent invested 100 billion RMB, CATL put in 50 billion RMB, NetEase, JD.com and IDG Capital each contributed 30 billion RMB, and the National Artificial Intelligence Industry Investment Fund invested 10 billion RMB.

Prior to this, Liang Wenfeng had put forward the principles of "no financing, no listing, no commercialization". This large-scale financing marks DeepSeek's official entry into the capital market, and has also sparked widespread industry attention to its commercialization path and technical vision.

At a recent investor exchange meeting, Liang Wenfeng elaborated in detail on DeepSeek's organizational culture, open-source logic, technical roadmap, and his views on the industry's competitive landscape.

The following is a collation of nearly 4 hours of Liang Wenfeng's speeches at the exchange meeting obtained by Tencent Technology, classified by theme, with a total of 118 entries. The text retains the original meaning as much as possible, with only minor edits.

01 Vision and Restraint

1. When we first started this company, our original intention was not to think about how much money we would make in the end, to go to the capital market, to go public, or anything like that. The first few dozen people never thought that way at all. If they had, they wouldn't have joined.
2. We came into this with enormous goodwill toward the world. We believe this is useful to humanity, something beyond money. Our original intention, our vision, and the vision we maintain to this day are not pursued in a way that maximizes commercial profits.
3. Managing a large company does not rely on rules and regulations, but on vision. Vision is not a slogan hung on the wall. Vision is how you act, not how you speak — it's your actual way of operating.
4. We are an organization without a rigid structure. We are vision-driven, organized entirely around a shared vision. We do not operate under the framework of "I need to hit certain KPIs, no assessments" — there is only the vision.
5. This vision is not even written down. Nothing about it has ever been documented. It exists in the way we do things and the attitude we take toward the world.
6. We don't have many other advantages. We don't have extraordinary capabilities. We are no richer than others, nor do we have better personnel than other companies — that's simply not the case. When we founded this company two years ago, we had little capital, few GPUs, little reputation, and no influence. We were just a group of very ordinary people.
7. The more restrained you are, the more likely you are to succeed — or at least this has proven true so far, and it's a logic that holds up. There's no other way to explain why we've made progress: we had no special advantages, a very low starting point, very few resources, and our team was just a random group of ordinary people.
8. The AI space is enormous, with massive potential benefits. We are very restrained. As long as we succeed in our mission, the final rewards will be extremely large. Even if we take only a small share, the returns will be substantial. So there is no need at all to think about how to carve out a piece of the pie right now, because the potential value is already big enough.
9. Last Spring Festival, our user base suddenly surged, but we did not try to retain those users at all costs, or monetize them immediately, or grab commercial benefits by cashing in on the user traffic. We didn't chase users or rush to make quick money, but we worked very hard to find ways to provide better services to our users.
10. We never had the idea that we want to build the next super App, compete with someone else, become the next ByteDance or the next Tencent — not at all. I believe the opportunities for AGI in the future will be enormous, and they will always be enormous.
11. Restraint is a strategy. Sometimes you can give up a few things in exchange for many more valuable things. The decision not to open source right now is the same — it can be seen as our self-imposed pressure, or as a way of giving back to the broader ecosystem.
12. In the long run, this kind of restraint can increase our probability of successfully building AGI. When considering any matter, I have no doubt that AGI will have tremendous commercial value. Based on that, my top priority is not how to grab a larger market share, but how to increase our chances of succeeding in the first place.
13. We have always been very restrained, unwilling to become rivals with any large or small internet company. I hope we can empower them, or assist everyone in this endeavor, and help everyone move forward.
14. I believe that by adhering to this attitude, we have not lost anything at all. We have not lost anything by open-sourcing, by acting with goodwill, or by offering help to others. On the contrary, it has given us an extra advantage. This seems counterintuitive, but it is indeed the case.
15. We are targeting AGI as our ultimate goal, but we have been doing commercialization all along — that's why we have C-end users and B-end revenue. From historical experience, this strategy has been successful.

02 AGI Roadmap

1. If you can describe a problem very clearly, provide it with complete context and instructions, the model already outperforms humans. But there is a premise here: you must give it complete context and complete instructions.
2. AI cannot replace your employees right now. But if AI gains the ability to learn continuously, just like your employees who spend two months learning on the job at the company, then it will be able to replace almost everyone. So we are just one step away from that next capability: continuous learning.
3. The development of AI can be understood as a series of steps. The step we took last year was Chain of Thought. We discovered that using the Chain of Thought approach could elevate intelligence to a higher level.
4. This year's step is Agent. We found that with the Agent approach, even more tasks can be accomplished, expanding the scope of its capabilities and raising the upper limit of its intelligence. Agents need to use CoT, and CoT builds on the previous step — the language model. None of these steps are wasted.
5. After Agents, we believe the next problem to solve is continuous learning: how to enable the model to learn continuously, instead of requiring heavy retraining. It should be able to engage in long-term, continuous learning just like a human being.
6. After continuous learning, we may reach a singularity. This singularity is the point where the model, after being able to learn continuously, can do everything humans can do. It will be able to develop its own new versions, conduct further research, build its own next iteration, and create better AI models.
7. This "singularity" is not an abrupt point — it is also a gradual process. This process may be a long, gradual change, not a sudden leap. But by convention, we all refer to it as a singularity.
8. This is our speculation. Our projected timeline is: first solve "learning to learn", then reach the intelligence singularity that enables self-iteration, and then move on to embodied intelligence. Once we reach embodied intelligence, AI will enter the real world, helping with housework and providing elderly care services.
9. If we solve continuous learning first, then solve the self-iterating singularity, then solve embodied intelligence, this journey will be very smooth. Because once you have the earlier capabilities, you can use that existing technology to help develop the next stage of technology.
10. We only focus on the main line toward AGI. The AI field is very broad, and there are many things we don't consider part of this main line — for example, 3D rendering and video generation. I don't think they have much to do with the core path of intelligence, so we won't work on them.
11. Video generation became very popular as soon as it came out, as if it was something you had to do, otherwise you wouldn't be considered a real AI company. I find that very strange. If you think about it carefully, it has almost nothing to do with the intelligence roadmap.
12. Commercially, it's a good business, yes. But it has nothing to do with the core of intelligence. We won't pursue something just because it's a good commercial opportunity. We only work on things that are part of the intelligence roadmap.
13. Based on our judgment, world models are not the most important priority at this current stage. The most important things are AI training, and solving the problem of continuous learning after AI training. This is our company's judgment — of course, every company may have a different view.
14. We now strongly believe in one narrative: AI can accelerate AI research. This means the process is not linear, because you can use AI to speed up your own research, leading to non-linear progress later on.
15. I believe we will eventually enter the embodied intelligence stage. Because for an ordinary person, their real needs are not computers, right? For normal people, their daily lives — food, drink, housing, transportation — don't revolve around computers. They need embodied intelligence to solve practical labor demands.
16. What do we want AGI to be able to do? We want it to help us iterate the next version of the model. Once we have embodied systems, we want them to help iterate the next generation of embodied hardware, to build the next version of robots.
17. The core capability of the next-generation model must be continuous learning — that's what defines it as the next generation. Before that, all we can do is optimize costs, improve performance, and increase speed. But to achieve a major breakthrough, the model must have continuous learning capabilities.
18. The current limited capabilities of Agents stem from their inability to learn continuously, to learn effectively over time. If we can complete the development of continuous learning first, AI's capabilities will be extremely powerful, and it will greatly enhance the efficiency of our own research work.
19. Once continuous learning is built first, general intelligence will be very easy to achieve, and using that capability will make everything much simpler. That's the outcome we hope for — it will save us a lot of effort. Otherwise, building general intelligence manually would be a very tiring, arduous task that is data-intensive and labor-intensive, with very low cost-effectiveness.

03 Team and Talent

1. Our past experiences have taught me that the AGI vision is extremely powerful. The talent advantage does not come from our people being smarter than others, but from how we organize those talented people, how we motivate them, and how we enable them to collaborate.
2. Gathering a group of smart people together does not automatically make them collaborate, or passionately work toward a shared goal. That's why you need a unifying vision.
3. Our top core priority is to maintain team stability. This is our most important core interest, and you could even say our only core interest. As long as we can keep the team stable, we will definitely succeed in building AGI — it's as simple as that.
4. Money is definitely not a problem, resources are not a problem, and all other factors are easy to obtain. For us, there is only one core priority, one non-negotiable principle: we must maintain team stability.
5. This is also a very big challenge we face, or I would say the biggest risk. Of course, this risk has been greatly alleviated by our recent financing round. Because the options granted to everyone are quite substantial, with very significant value.
6. In terms of team stability, as long as our most important, earliest employees remain stable, the rest of the team is unlikely to leave. Even if others get fewer options or lower salaries, they won't leave. Because they didn't join just for money. Everyone wants to work in an environment that can actually deliver AGI.
7. Everything else is just a matter of time. At worst, it might delay us by half a year or a year, but we will definitely deliver results. We definitely don't lack capital, we definitely don't lack resources — none of those are issues.
8. The gap between us and the United States lies mainly in resources. The gap in talent is not large. There is almost no talent gap, because it's the same pool of people — many of them are Chinese. When Chinese people go overseas, some stay in China, some stay abroad. It's not that all the smart people left the country — that's not true at all.
9. Talent is not the bottleneck. Resources are the biggest bottleneck. Resource constraints first affect talent development, because with less computing power, we can run fewer experiments, so our talent pool as a whole lags behind the US. The talent gap essentially comes down to the computing power gap.
10. The shortage of AI talent is also a temporary problem, and we have already seen it greatly alleviated. There is no real shortage of AI professionals — every company can quickly train their own people, and talent cultivation is very fast.
11. There are currently too many companies in China building foundation models, far too many. The US might only have three major players, but China has far too many teams working on foundation models. Eventually, we won't need so many people doing this work, and the industry will naturally consolidate.
12. Our company's management actually operates on two tracks: one is top-down, the other is bottom-up. The bottom-up approach means everyone can decide what they want to work on, no one is supervising them, and there are no KPIs.
13. We generally hope that employees have half of their time unassigned, free to work on whatever they want. This is the research zone, allowing them to explore independently, focusing on whatever they think is important, with no pre-set requirements.
14. We generally don't advocate working overtime. There are two reasons for this. First, research requires a relaxed environment. If you push people too hard, you can't do good research. Because research requires personal curiosity and continuous thinking, it can only happen in a relaxed atmosphere that allows for exploration.
15. Second, we are extremely focused. Being extremely focused means we have very few things to do. So we don't have that much work piling up, and there's no need to work overtime. This is fully consistent with the principle of restraint we talked about earlier.
16. Our entire company is built on consensus. I don't make all the decisions alone — I seek consensus. My authority and influence within the company are based on broad consensus from the team.
17. This decision-making mechanism is essentially a consensus-seeking process. I can't push any initiative forward unless it has broad consensus, and I will only promote it once that consensus is reached.
18. As our team grows, we will make corresponding adjustments. In fact, we need to make these adjustments immediately, and we are already in the process of doing so. Without these adjustments, many things will be impossible to move forward. There are indeed many departments that need formal organizational structures.

04 Computing Power and Resources

1. How many GPUs do we need? Right now, the more the better, without a doubt. As long as it's within our affordable range, more GPUs are always better. So our current strategy is to buy as many GPUs as we can at a reasonable price point.
2. In reality, it's extremely difficult to spend that huge sum of money. You can't get that many GPUs — they're very hard to purchase, prices are very high, and we can't overpay excessively. If we can spend 20 billion RMB this year, our procurement team will have delivered an absolutely outstanding performance.
3. The biggest gap between us and the United States is in resources. On one hand, we simply can't get enough GPUs in the domestic market. On the other hand, our capital investment is far less than that of the US. Our capital investment is much lower overall, and talent salaries account for a tiny fraction of total costs. You see US companies offering salaries up to 100 million USD, but even then, personnel costs are a small share — the vast majority of costs go to computing power.
4. All the differences we observe — differences in talent, model capabilities, and application performance — can be attributed to the gap in computing power resources.
5. The gap between us and the US might be 12 months, 12 to 18 months, or 6 to 12 months. In short, we are roughly two years behind the US, but we have achieved what we have with only 1/20 of their computing power.
6. The current narrative is that we are 1-2 years behind, but using only 1/20 of their computing power. Our goal is to rewrite that narrative: in the future, we will use a fraction of their computing power, but shorten the time gap to 6 months, or even 3 months. That's the target we're aiming for.
7. We believe in Scaling Laws. The larger the scale, the better the performance, and the more capabilities we can unlock. What's stopping us from scaling is purely a lack of computing power — not that we don't want to scale, but that we don't have enough computing resources to do so.
8. The reason we train such large models is not that we think this size is sufficient, but that it's the maximum we can afford given our available resources. We calculate the maximum model size we can support based on our resources — that's how we determine it, not that this model is already perfect.
9. When Silicon Valley says that Sc