HomeArticle

What on earth is so distinctive about Liang Wenfeng's way of thinking?

砺石商业评论2026-08-21 13:18
Liang Wenfeng's unwavering long-termism: focus on AGI, no rush for monetization

When you string together Liang Wenfeng's organizational philosophy, you will see a complete closed loop: he does not pursue immediate commercial returns, so he can give employees maximum freedom; free exploration brings real creativity, creativity consolidates the intergenerational advantage of technology, and technological advantages are finally transformed into commercial returns. Trade freedom for creativity, and trade creativity for the future. This is the most complete and self-consistent part of Liang Wenfeng's thinking model.

On May 20, 2026, a 3-hour and 44-minute closed-door investor meeting was held in Hangzhou. The outside world learned of the existence of this meeting through a 42-page audio transcript — it first circulated in a small circle of investors, then was made public, and sparked public opinion because its content was too "unconventional".

Almost every sentence Liang Wenfeng said at this meeting challenged the mainstream logic of China's technology industry. He said "DeepSeek has only one main line, which is AGI, and everything else is a strategy"; he said "The more restrained you are, the more likely you are to make this happen"; he said "I don't want to be the next ByteDance or Tencent, and we don't take commercialization as our goal"; he said "If your vision is to take as much as possible, you lose first". Finally, he even said: "The narrative I like is 'a group of ordinary people have done extraordinary things'."

This meeting minute is probably the most important, complete and timely public text for understanding Liang Wenfeng's thinking model. It was born at a special moment: a few days earlier, DeepSeek had just completed the largest single round of financing in the history of China's AI industry — about 51 billion yuan, with a post-investment valuation of about 400 billion yuan, and Tencent, CATL, NetEase, JD, and the National Artificial Intelligence Fund all participated. At this capital feast where "money is too hot to handle", what Liang Wenfeng repeatedly emphasized to the investors who paid real money under the stage was not growth, not monetization, not returns, but restraint.

Putting this 2026 collection of ideas together with his only two in-depth interviews in 2023 and 2024, an astonishing fact emerges: three years have passed, the landscape of China's AI industry has been turned upside down, but Liang Wenfeng's core judgment has hardly changed, and it has only been sharpened. This "unwavering stance for three years" itself is the most distinctive feature of his thinking model.

Restraint is a strategy: active abandonment of trivial opportunities

The first keyword of Liang Wenfeng's thinking model is "restraint". And this "restraint" is not conservatism, but a deeper level of enterprising spirit.

At the 2026 closed-door meeting, he said a passage that can be regarded as the general outline of DeepSeek: "DeepSeek has only one main line, which is AGI. Everything else is a strategy." "The more restrained you are, the more likely you are to make this happen."

The key to this sentence lies in his redefinition of the relationship between "restraint" and "achievement". In mainstream cognition, restraint usually means giving up and gaining less. But Liang Wenfeng turns restraint into a strategy — giving up immediate opportunities is to increase the probability of achieving the ultimate goal.

This kind of "restraint" is almost resolute when implemented in specific choices. He explicitly stated that he would not do 3D, video generation, or world models. These are not "unachievable". On the contrary, in 2026, video generation and world models are the most sought-after tracks by capital. A team with DeepSeek's technical base can cut into these tracks with an almost dimensionality reduction strike. But Liang Wenfeng said bluntly at the closed-door meeting: "I don't want to be the next ByteDance or Tencent."

He even set a boundary for "commercialization" itself. When talking about API pricing, he said: we do not seek exorbitant profits, only reasonable profits. The pricing standard is calculated based on recovering equipment costs in ten months and a profit of about six times. Further price reduction will not bring demand elasticity and makes no sense.

This "restraint" is in the same line as what he said in the 2024 interview — "Grabbing users is not our main purpose" "We just do things at our own pace" — but in 2026, it becomes clearer and more aggressive.

It is necessary to understand the real logic behind Liang Wenfeng's "restraint". He said a seemingly contradictory sentence as early as 2023: "If you must find a commercial reason for it, you may not find it, because it is not cost-effective. From a commercial point of view, the return on investment of basic research is very low." He is not ignorant of business. On the contrary, he understands the priority of business too well — at the moment of technological intergenerational change, putting resources on the "right thing" rather than the "profitable thing" will bring higher long-term returns.

In his words in 2026: AGI may eventually account for 10% of human GDP. "As long as it is successfully developed, even a small share of the benefits will be very huge." So he is not in a hurry to grab market share, because he judges that the decisive factor of this game is "the probability of success", not "the share obtained". This is not lofty aloofness, but an extremely shrewd account.

The only main line AGI: turning the ultimate goal into a specific technical route

If "restraint" is the surface of Liang Wenfeng's strategy, then "persistence in AGI" is its core. What makes him different is that he breaks down the grand concept of AGI into an extremely specific and clear technical route.

At the 2026 closed-door meeting, he gave a clear intelligence development ladder: Language Model → CoT (Chain of Thought) → Agent → Continuous Learning → Self-iteration Singularity → Embodied Intelligence, and clearly pointed out that the next key bottleneck is "Continuous Learning" — a problem that has not yet been solved by the global academic community. "AI now does not lack taste and intuition, but the ability to continuously learn." He also said that multimodality and search are only "components", not the main line of intelligence itself.

This passage contains a huge amount of information. It shows that Liang Wenfeng's pursuit of AGI is not a vague belief, but a combat map with a clear roadmap, milestones and priorities. He knows which stage he is at (Agent), where the next step is stuck (Continuous Learning), and which paths are detours (multimodality, world models).

This ability to "materialize grand goals" was reflected as early as his first interview in 2023. At that time, he said: "Our destination is AGI, which means we need to study new model structures to achieve stronger model capabilities with limited resources." He even bet on three directions in 2024 — mathematics and code, multimodality, and natural language itself, on the grounds that "mathematics and code are the natural test fields for AGI, a bit like Go, which is a closed and verifiable system".

DeepSeek's technological progress in 2026 proves that this route is not empty talk.

On April 24, 2026, DeepSeek-V4 was released — exactly the same day that OpenAI released its new model. It continues the MoE architecture, with the core of million-token context. Through the hybrid mechanism of Compressed Sparse Attention (CSA) and Heavy Compressed Attention (HCA), the attention calculation complexity is significantly reduced from O(n²). V4 has a comprehensive score of 98.5, surpassing GPT-5.4 on SuperCLUE; its Codeforces programming score is 3206, reaching the top 0.01% of human programmers; by May, the monthly token usage of V4-Flash reached 18.4 trillion, surpassing GPT-4o to become the world's most widely used large model.

The most notable point is: V4 is the first cutting-edge Chinese large model that runs on domestic computing power at its debut — the underlying code is fully migrated to the Huawei CANN ecosystem and runs on the Ascend 950PR chip.

Liang Wenfeng calculated an account at the closed-door meeting: DeepSeek currently has about 20,000 H-equivalent computing cards, using about 1/20 of the computing power of its US counterparts, and the model is 1-2 years behind. But he judged that "there is almost no gap in talent — it is the same group of people", and the real gap is "resources (cards), not talent". His goal is to shorten the Sino-US gap to 3-6 months with a higher proportion of computing power investment.

Behind this passage is an extremely clear confidence: he does not think that China's AI industry lags behind in intelligence, but only in resources; and the resource gap can be made up for by strategy and efficiency. This judgment allows him to set the goal of "catching up and surpassing" even when the computing power is only 1/20 of his competitors.

Commercialization: Do not pursue excessive monetization, and believe that returns will come naturally after technological differentiation

On the topic of "commercialization comes naturally", although Liang Wenfeng has not said these exact four words, all his statements in 2026 almost interpret this logic.

At the closed-door meeting, he said: "We have actually been doing commercialization all the time, but commercialization is not our goal. The time point for DeepSeek to fully shift to commercialization may be very far away." "We only pursue reasonable profits, and will not set prices at the level of profit maximization." "Next year, after B-end revenue expands, it is expected to approach net profit. In the worst case, selling APIs alone is enough to support a listed company."

This passage answers an unavoidable question for all AI companies: how to survive without commercialization? Liang Wenfeng's answer is twofold. First, DeepSeek is indeed doing commercialization, with an annualized recurring revenue of about 400-500 million US dollars, mainly from API calls, and a gross profit margin of more than 50% — it has no shortage of cash flow. Second, the pace of commercialization is completely subject to the progress of AGI, not the other way around.

His ranking of competition order reveals the underlying logic of his business thinking: "Competition order: cost first, time second, user experience third." This is a very counter-intuitive ranking. The vast majority of Internet companies put "user experience" first. But Liang Wenfeng puts "cost" first, because he judges that the large AI model is essentially infrastructure — and the core of infrastructure competition is low cost. "The large model is essentially infrastructure, and the core is low cost." This sentence is the condensation of his business view.

This "cost-first" idea allows DeepSeek to take a different path from all its competitors. It does not rely on subsidies to grab users, but on the cost advantage brought by the ultimate technology. V4-Flash outputs 0.07 dollars per million tokens, which is only 1/35 of GPT-4o; the V4 open source model is completely consistent with its own deployment model, "we will not open source a worse model and use a better one ourselves".

In an era when almost everyone is anxious about "how to monetize large models", Liang Wenfeng is one of the very few people who truly believe that "make technology to the extreme first, and commercial returns are by-products". The foundation of this confidence is his certainty of technological differentiation: when your technology forms an intergenerational advantage that others cannot catch up with in a short time, the commercial path will naturally emerge.

Organization and talent: give employees freedom, and trade freedom for creativity

The most counter-intuitive and distinctive part of Liang Wenfeng's thinking model is organizational management.

At the 2026 closed-door meeting, he brought DeepSeek's organizational philosophy to the extreme. He said: "The only core interest of the company is team stability — as long as the team does not fall apart, AGI will definitely be achieved. This is the core purpose of this financing."

This sentence must be understood in context. In 2025 before the financing, many core backbones of DeepSeek were poached by ByteDance, Tencent, Xiaomi and other companies, including the core R1 researcher. The primary purpose of Liang Wenfeng's financing this time is not to buy computing power, but to give core researchers a credible valuation, make their options meaningful, and thus retain the team. In the eyes of others, this is a capital operation; in Liang Wenfeng's eyes, it is the capitalized expression of the only core interest of "team stability".

His definition of organization is also very subversive: "DeepSeek is a vision-driven organization, not a KPI-driven organization, and the vision is not even written down." "The decision-making mechanism is based on consensus, not personal authority." "Spend half of the time doing top-down key work, and the other half of the time for free exploration." "We generally don't work overtime, and research requires a relatively relaxed environment."

"No KPI, unwritten vision, consensus-based decision-making, no overtime" — he expressed this set of management philosophy as early as 2023 ("We have no KPI, no so-called tasks" "Innovation requires as little intervention and management as possible"), and also in 2024 ("Everyone can call the cards of the training cluster at any time without approval"). But in 2026, he tied it to the judgment that "team stability is the only core interest", giving this set of philosophy a harder core.

His employment philosophy is also consistent. In 2024, he said: "There are no unfathomable wizards, all are fresh graduates from top universities, fourth and fifth year PhD interns who have not graduated, and some young people who have only graduated for a few years." He prefers young people, because "innovation first requires confidence, which is usually more obvious in young people"; he does not use "top talents" returning from overseas, because "the top 50 top talents may not be in China, but maybe we can cultivate such people ourselves".

The greatest trust he gives to young people is freedom. This freedom brings real creativity — DeepSeek's self-developed MLA architecture (Multi-head Latent Attention), TileLang compiler, and Compressed Sparse Attention, these industry-changing technologies, according to his description, were mostly born from the free exploration of young researchers, rather than top-down task assignments.

When you string together Liang Wenfeng's organizational philosophy, you will see a complete closed loop: he does not pursue immediate commercial returns, so he can give employees maximum freedom; free exploration brings real creativity, creativity consolidates the intergenerational advantage of technology, and technological advantages are finally transformed into commercial returns. Trade freedom for creativity, and trade creativity for the future. This is the most complete and self-consistent part of Liang Wenfeng's thinking model.

Vision: China's AI industry cannot always follow others, and open source is giving away benefits

The last dimension of Liang Wenfeng's thinking model is to incorporate "China" and "open source" into his own technological narrative. This is also where he differs from most Chinese entrepreneurs, and even from Silicon Valley AI leaders.

As early as 2024, he put forward the famous judgment: "China's AI industry cannot always be in a following position. The real gap is the difference between original creation and imitation." At the 2026 closed-door meeting, he upgraded this idea more thoroughly, and directly talked about "open source is giving away benefits":

"Open source is giving away benefits: internally, employees have a sense of achievement, and the company has cohesion; it is also good for society." "I have no doubt that AGI will have great commercial value. On this basis, my priority is not to get more share, but to increase the probability of success." "The open source models we provide are exactly the same as the models we deploy ourselves. We will not open source a worse model and use a better one ourselves."

This statement hides a major thinking change. In 2024, he called open source a "cultural behavior" ("Open source is more like a cultural behavior than a commercial behavior, and giving is actually an extra honor"); in 2026, he went a step further and elevated open source to a strategic height — because AGI is "big enough that no one can achieve it alone", open source is not passive giving away, but an active strategy chosen to increase the probability of "achieving AGI".

His definition of open source almost overthrows the common sense of the business world: "In the face of disruptive technologies, the moat formed by closed source is short-lived. Even if OpenAI keeps its model closed, it cannot prevent being overtaken by others. So we precipitate value on the team... forming an innovative organization and culture is our moat."

Behind this is a redefinition of the concept of "moat". Most companies regard moats as technical secrets, patents, and data barriers; Liang Wenfeng believes that there are no secrets in technology, "it only takes time and cost to reset", and the real moat is the team, organization and culture. So he dares to open source, because he judges that the moat is not in the model weights at all.

He even made a specific judgment on the future of China's AI industry. When talking about the computing power gap, he said that domestic chips have a "historic opportunity", because "AI + advanced compiler (self-developed TileLang) is rapidly disintegrating NVIDIA's CUDA moat", and Huawei