The AI battle is in full swing, and everyone is waiting for Liang Wenfeng.

Liang Wenfeng

Author/Feng Yuchen

At this moment, the gods have taken their positions.

Internet giants are scrambling to get to the table, and their ambitions are clear: they all want their own AI large models to become the strongest traffic entry points. In the arena of AI large models, that man was once the protagonist.

Exactly one year ago, he rose to fame. Going back to the week before the Spring Festival in 2025, Liang Wenfeng led DeepSeek to release the large inference model R1. With unexpectedly low costs, it achieved performance comparable to that of the world's top models. Since then, the mysterious power from the East has successfully conquered the global tech circle.

This post - '85s guy from Zhanjiang, Guangdong, rarely appears in public and has a net worth of over billions. The status of DeepSeek, the company he founded, needs no further introduction. However, during the fierce battle of Chinese AI large models, DeepSeek seems to have gone into hiding. On February 11th, DeepSeek quietly updated and upgraded. When the outside world speculated, "Is the new version of DeepSeek coming?", Liang Wenfeng and DeepSeek remained silent.

Sometimes, the calmer one is, the more afraid the opponents will be.

In the heat of the AI battle, Liang Wenfeng stays quiet

This is a very delicate moment in the history of Chinese AI.

Let's first look at the most tangible thing for the public - the cash giveaways. At the end of January, Tencent distributed 1 billion yuan in cash red envelopes to incentivize users, and at the same time, Baidu invested 500 million yuan in cash red envelopes to promote its Wenxin Assistant. Soon, Alibaba's Qianwen came in with 3 billion yuan in red envelopes, which was like a bombshell in the milk tea shop business. ByteDance, the so - called "universe factory", also took action. Users participating in the Spring Festival lottery on the Doubao APP could win up to 8,888 yuan in cash.

With funds burning and users in ecstasy, it reflects the game of the future fates of the top AI large models.

On the product side, ByteDance set off a big fire. On February 14th, ByteDance officially announced the release of the Doubao large model 2.0. And on February 7th before that, ByteDance's video model Seedance 2.0 went viral across the Internet. Netizens exclaimed, "The moment when a renaissance and a technological revolution happen simultaneously has arrived." Almost at the same time, Alibaba launched a new generation of image - generation basic model Qwen - Image 2.0. Zhipu and MiniMax also joined in and released new models together.

This kind of "responding to every move" shows a rather strange competitive tacit understanding. And when you look back suddenly, DeepSeek, the one that once sparked the large model open - source revolution, seems rather calm.

Actually, some news leaked out earlier.

According to foreign media The Information, DeepSeek planned to release a new generation of flagship artificial intelligence model codenamed V4 around the Spring Festival in mid - February, aiming for the throne of Coding AI. But soon, there were reports saying that V4 might not be released on time, and it was mentioned that DeepSeek remained silent about the release time.

On February 11th, DeepSeek quietly updated, increasing the context window length from the previous 128K tokens directly to 1M (one million) tokens. When the outside world speculated that this version might be the V4 Preview, the official still remained as silent as ever.

But this mysterious dark horse that "doesn't care about the battle" always attracts the attention of the outside world.

In January 2026, the DeepSeek team published two papers signed by Liang Wenfeng - mHC (Manifold - Constrained Hyper - Connection) and Engram (Conditional Memory). At that time, the market began to speculate: Is this related to the core technology of DeepSeek V4?

During the Chinese New Year period when many companies slow down their recruitment, DeepSeek is still continuously recruiting on various platforms. For example, on the official account of DeepSeek, on February 5th, it updated recruitment information for multiple positions such as product managers and client - side R & D engineers.

"Most of our companies are used to following rather than innovating," Liang Wenfeng once said. Now, in the chaos of the "gods' battle", the outside world is looking forward to Liang Wenfeng, the disruptor who once created miracles through innovation.

The county top - scorer and the Chinese DeepSeek moment

At the southernmost tip of the Chinese mainland, the Mililing Village in Zhanjiang, Guangdong, was once bustling with tourists.

Liang Wenfeng, born in 1985, is from here. His parents are both local primary school teachers. He has been an excellent student since childhood. He finished high - school mathematics in junior high school and even started learning college - level mathematics. But rather than exam rankings, "what problems were solved today" was what his parents cared more about during Liang Wenfeng's teenage years.

This kind of growth background complements the fact that he is later regarded as a "person with a pure technological ideal" in the AI circle many years later.

At the age of 17, as the top scorer in the college entrance examination from Wuchuan No. 1 Middle School, Liang Wenfeng was admitted to the major of Electronic Information Engineering at Zhejiang University. In 2007, he continued his postgraduate studies in the major of Information and Communication Engineering at Zhejiang University, under the supervision of Xiang Zhiyu, mainly engaged in machine - vision research.

Actually, at that time, computer science was a rather niche major. Civil engineering, electrical automation, finance and international trade were more popular choices. There was a stereotype that those who studied computer science would end up repairing and selling computers in computer malls after graduation, "just being a wage - earner".

But Liang Wenfeng chose this path. Personal interest and the great changes of the era laid the groundwork for the future of this genius teenager.

In 2008, the global financial crisis swept across the world. At that time, as a student, Liang Wenfeng keenly captured the opportunities behind the crisis. He formed a team with his classmates and started exploring fully - automated quantitative trading using technologies such as machine learning. In 2015, Liang Wenfeng and his Zhejiang University classmate Xu Jin officially founded Hangzhou Magic Square Technology Co., Ltd. - which later became the well - known Magic Square Quantitative.

AI has run through Magic Square from beginning to end. In the early days, almost all quantitative strategies of Magic Square used AI models for calculation. In 2019 and 2021, Magic Square invested over 1 billion yuan to build its self - developed deep - learning training platforms "Firefly One" and "Firefly Two".

Going forward smoothly, by 2021, Magic Square Quantitative successfully entered the ranks of 100 - billion - level quantitative private funds. In this year, Magic Square Quantitative obtained NVIDIA A100 graphics cards, and the number of graphics cards installed in its AI cluster reached the "tens of thousands" level. It's worth noting that at this time, there were only a few companies in China with more than 10,000 GPUs. Except for Magic Square Quantitative, the rest were almost all Internet giants.

At the age of 38, Liang Wenfeng, the financial tycoon, made a decision: to found DeepSeek and develop general artificial intelligence.

In his eyes, this might be a part of satisfying his curiosity about the technological world. What DeepSeek wants to do is not just to replicate the already - famous GPT. Instead, just like its company name, it wants to seek deeply - truly conduct research on large models and lead the industry ecosystem.

The vision soon became a reality, and the debut was extremely shocking. In January 2025, one week before the Spring Festival, DeepSeek, with a nearly 100 - person team, launched the DeepSeek R1 version and directly opened the source generously. Its inference ability was comparable to the world's top - level models, but the cost was only about several hundred thousand dollars, completely subverting the perception that top - level models require tens of millions of dollars in investment.

Thus, Liang Wenfeng left a precious mark in the history of the Chinese AI process: the DeepSeek moment.

The battle of the gods and a new era of Chinese AI

"AI is still meant to be used by people", and this is even more urgent now.

Just like DeepSeek, which always pursues efficiency and is dedicated to research, it has also started to focus on specific users at the consumer end. In a recent recruitment for a product - manager position, the responsibilities section reads:

· Lead the functional planning and experience design of DeepSeek's consumer - end products, keenly perceive user needs, and continuously promote product innovation and experience optimization;

· Identify and break down problems, reduce user usage costs, focus on user growth, and improve usage depth, stickiness, and conversion;

·......

Over the past year or so, many consumer - end users of DeepSeek once suspected that DeepSeek didn't have a product manager. "The pages and entrances are very simple, as if it doesn't need to make money," a personal user of DeepSeek said, and even joked: It has finally realized that we consumer - end users are a key part of achieving AGI.

Compared with DeepSeek's modesty and reserve, the big companies that are more eager for commercialization keep coming up with new strategies.

Just like Alibaba spending 3 billion yuan on red envelopes to let Qianwen order milk tea. It may seem like burning money to buy traffic, but in fact, it's a necessary step to cultivate users' awareness in the AI era. Through the subsidy method of free orders, which is very familiar to the public, it first attracts ordinary people to use Qianwen, and then tries to make them realize that AI can place orders and even "do things".

Similarly, ByteDance's Doubao, which is going to give gifts during the Spring Festival Gala, has already found a foothold in many life scenarios: guiding people on Korean - style handsome - guy dressing, taking artistic photos of pets, tutoring children with their homework... In these role - playing and emotional conversations, Doubao's monthly active users have already exceeded 200 million.

All these indicate that in the post - era of large - model competition, the explosion point of demand for AI consumer - end applications is approaching.

This also brings the competition back to the most fundamental essence: whoever can solve users' real needs and problems can become the super - entry point in the AI era.

For the Internet giants that always need traffic, this battle is of even greater significance - the opportunity to define the Internet order for the next decade is right in front of them.

This article is from the WeChat official account "Investment Circle". Author: Feng Yuchen. Republished by 36Kr with permission.