Liang Wenfeng and Yao Shunyu, submit your papers in April.

Written by | Huahua

In April 2026, the Chinese AI circle will witness a rare head - to - head competition.

DeepSeek V4, a multi - modal large - scale model that Liang Wenfeng has polished for a long time, is planned to be officially launched.

Almost at the same time, Tencent's new Hunyuan model will also be released. The person in charge is Yao Shunyu, who just returned to China at the end of 2025 and took office with the title of Chief AI Scientist of Tencent's Executive Committee.

Two companies, one in the south and one in the north, one a startup and one a giant.

One is a science - and - engineering guy who has disrupted the AI industry with a quantitative private - equity mindset, and the other is an academic figure invited back from overseas by Tencent with the highest - level treatment.

It's almost impossible for them to be at the same table, but in April, they both have to submit their answers.

This is not a coincidence. It's a turning point.

1. Tencent's Decision

To understand what April means, we must go back to the decision Tencent made in early 2025.

At that time, in the domestic large - scale model market, Baidu had Wenxin, Alibaba had Tongyi, and ByteDance had Doubao. Each company was vying for market share. Tencent's Hunyuan was not in the first echelon, and its product Yuanyuan hardly had any presence.

At this time, Tencent chose to integrate DeepSeek into Yuanyuan.

Many people interpreted this decision as a practical move at that time. Instead of insisting on self - research, it's better to use the best model to retain users.

But it was also an admission: in terms of large - scale model capabilities, we are currently inferior to a startup.

The data of Yuanyuan verified the correctness of this decision.

After integrating DeepSeek, the number of users and user activity significantly rebounded. Tencent exchanged traffic for time, using Yuanyuan's existing product ecosystem and the traffic - guiding capabilities of WeChat and QQ to buy a window period for Yao Shunyu's team to develop a self - developed model.

However, the time window is ultimately limited. Once the latest version of Hunyuan is launched, the first real question Tencent faces is: which name do the users remaining in Yuanyuan actually recognize?

This is not a question that can be answered by market research. It can only be answered by the retention data after April.

2. Another Challenge for DeepSeek

DeepSeek is facing a completely different kind of pressure.

After the release of V3, DeepSeek's figures were as follows: From its launch to February 9, 2025, the cumulative downloads exceeded 110 million, and the weekly active users reached a peak of nearly 97 million.

This scale was close to Twitter's daily active users at its peak. For an AI application launched less than half a year ago, there were hardly any precedents at that time.

A greater impact occurred at the industry level. After the release of V3, Nvidia's market value evaporated by about $600 billion in a single day, and the US tech circle began to reassess the logic of the computing - power arms race.

DeepSeek used an open - source model to challenge the industry consensus that large - scale models must rely on brute - force computing power in terms of both cost and efficiency.

But these figures and this narrative have also become the burden that V4 must bear.

The market's expectation for V4 is not just to be better than V3, but to redefine the standard again. This is a very difficult task because V3 has already set a very high bar.

Liang Wenfeng's approach to dealing with this is to make the iteration direction of V4 specific, rather than continuing to advocate for an efficiency revolution.

Judging from the pace of academic papers, this preparation started long ago.

In December 2025, "mHC: Manifold - Constrained Hyper - Connections" co - authored by Liang Wenfeng's team was published, focusing on optimizing the connection method of the underlying architecture.

In January 2026, "Conditional Memory via Scalable Lookup" signed by Liang Wenfeng proposed a conditional memory mechanism. During the inference process, the model can dynamically retrieve and activate relevant memories according to conditions, rather than stuffing all historical contexts into a fixed window.

These two papers were published only a few months before the release of V4. They are not just academic demonstrations but more like technical endorsements for the product roadmap.

Another aspect of V4 is that it is being specifically optimized for domestic chips and is expected to become the first top - tier large - scale model to run entirely on a domestic computing - power ecosystem.

Previously, the training and inference of all top - tier large - scale models were highly dependent on Nvidia's GPUs. Domestic chips such as Huawei Ascend and Cambricon still lag behind Nvidia in terms of performance and software ecosystem, and the adaptation work is far more complex than it sounds.

If V4 can successfully run on domestic chips, it means that for the first time, there is a verifiable alternative path to address the computing - power dependency, which is the most critical link in China's large - scale model industry chain.

3. Two Paths, One Destination

Interestingly, DeepSeek and Tencent are converging in the same direction in terms of technical routes: long - context, long - term memory, and Agent usability.

This is not a coincidence. It is the consensus being formed across the industry. The core of the next - generation AI competition is no longer who has more parameters, but who can truly remember users, understand context, and work continuously on complex tasks.

However, the paths taken by the two companies to reach this destination are completely different.

DeepSeek's approach is to make fundamental changes to the underlying architecture. The conditional memory mechanism aims to fundamentally change the model's memory method.

Tencent's approach is to first define evaluation criteria. CL - bench has proposed a new benchmark for measuring context - learning ability.

If this set of standards is accepted by the academic circle and the industry, Tencent will gain an initiative in narrative: whether your model is good or not will be judged by the standards I set.

These two paths represent two different competitive strategies. Making a better product and defining what a better product is are two different things.

Historically, the latter sometimes lasts longer. But the prerequisite is that the standards you set must be accepted by others. If the capabilities of DeepSeek V4 directly surpass the evaluation dimensions set by CL - bench, then those standards will just be Tencent's self - assessment.

4. After April

April will come, the two "examination papers" will be opened, and the market will make its judgment.

But the real questions worth asking are far more complex than whose model is better.

Question 1: Can Tencent handle the user transition?

If Tencent can successfully transition from integrating DeepSeek to launching its self - developed Hunyuan, considering Tencent's scale, data, and access to various scenarios such as WeChat, QQ, Tencent Video, games, and Enterprise WeChat, the value of these scenarios can be truly unleashed once there is a sufficient underlying model.

However, if user retention declines after the transition, Tencent may need to re - evaluate its position in this competition.

Question 2: Can DeepSeek maintain its position as the number one domestic model?

After V3, DeepSeek is no longer just a product. It is a standard and a narrative.

V4 needs to prove that this is not a coincidence but a sustainable ability.

Especially in the aspect of domestic chips, if it succeeds, DeepSeek will have a unique narrative that other top - tier large - scale models do not have, and the value of this narrative may even be greater than the model itself.

Question 3, and the biggest one: Who is competing for the right to define?

In the second half of the large - scale model era, it's not just about whose capabilities are stronger, but also about who gets to define what the next - generation AI should be like.

Liang Wenfeng's answer is: Open - source, efficient, with strong memory, and running on its own chips.

Yao Shunyu and Tencent's answer is: Deep integration with scenarios, long - context, and Agent - friendly. These two answers are not mutually exclusive, but they represent two different paths, one starting from the model and moving towards applications, and the other starting from applications and moving towards the model.

April is just a turning point. The real answers may take the whole year of 2026 to gradually emerge.

But one thing is certain. In early 2025, when Yuanyuan integrated DeepSeek, it was a reshuffle of the Chinese AI landscape. For the first time, a large company publicly admitted that a startup had taken the lead in core technology.

If Hunyuan and V4 are launched simultaneously in April 2026, it will mark the beginning of another reshuffle. It's no longer about who is in the lead, but about which path each company will take.

The story of Chinese AI is evolving from a race to a divergence of paths.

And divergence is often more interesting than a race.

Words from [Beyond the Page]:

Two companies, two paths, one time point.

What's really interesting is not the scores. It's that in these two "answersheets", we will see two completely different senses of security.

DeepSeek's sense of security comes from making things happen, with an open - source, efficient model running on its own chips. This approach can be sustainable.

Tencent's sense of security comes from integrating things into its existing platforms, such as WeChat, QQ, and games.

For Tencent, submitting the "answer" in April is not about proving something to the outside world. In essence, it's about "redeeming" Yuanyuan's "brain" from DeepSeek.

This article is from the WeChat official account "Beyond the Page" , author: Huahua. Republished by 36Kr with permission.