GuShiio.com Wen| DataEye
On January 20, just a few hours apart, Deep Quest released DeepSeek-R1 and Kimi K1.5 multimodal thinking models with similar technical paths to the Dark Side of the Moon.
Later, OpenAI stated in a paper that the two China AI companies simultaneously independently discovered the secret of o1 (OpenAI o1 Inference Model).
But judging from the results, the two are completely different. The former is famous and rewriting the global AI landscape, while the latter is not too loud.
Prior to this, no one would have thought that an AI model incubated by a hedge fund company without any advertising could break the OpenAI myth in just a month, and even made netizens shout: DeepSeek short the United States!
How popular is DeepSeek? You can learn one or two from the data:
The WeChat index shows that in the past 30 days, the average WeChat index of Doubao and Kimi has been around 30 million and 21 million respectively, while the average value of DeepSeek’s WeChat index has reached 590 million, with a peak of nearly 1 billion, which is dozens of times that of Doubao and Kimi.
At this point, the entire Internet has shifted from All in AI to All in DeepSeek.
So, DeepSeek-R1 has been born for one month and has been popular ever since. How are Doubao, Kimi, and Wen Xiaoyan now?
1. Embracing faction: Tencent Yuanbao, Wen Xiaoyan
The emergence of DeepSeek has further differentiated major Internet companies such as BAT and the AI Six Dragons such as the Dark Side of the Moon.Or embrace DeepSeek, seize advantages and craze to expand territory, or rely on your own and explore more possibilities for AGI formats.
Among them, there are not a few who embrace DeepSeek, and Tencent is one of them.
On February 16, Tencent confirmed that WeChat had launched the AI search function and officially grey-tested the DeepSeek-R1 model to provide in-depth thinking services.
As a terminal application for Tencent’s large model, Tencent Yuanbao quietly updated on February 13, announcing access to DeepSeek; on February 17, Tencent’s self-developed deep thinking large model HunyuanT1 also officially launched the Tencent Yuanbao APP for grayscale testing.
At the same time, Tencent Yuanbao also increased its launch efforts at almost the same time, seemingly to help with the new versions of DeepSeek and HunyuanT1.
ADX industry data shows that from February 1st to 14th, Tencent Yuanbao’s release intensity has remained relatively low, with an average daily volume of materials released about 2000 groups. Since February 15th, the release intensity has increased rapidly, and the daily volume of materials released exceeds 3400 groups. As of February 18, Tencent Yuanbao’s daily release volume reached a new high of 6570 groups.
Although connecting the DeepSeek model to its own AI assistant will make people feel a bit like doing a wedding for others, Tencent doesn’t seem to care. On the contrary, judging from the current situation, it includes Yuanbao, WeChat, Tencent Docs, QQ Browser, QQ Music, Tencent Maps and other applications, at least a dozen of Tencent’s products have announced access to the DeepSeek model.
According to DataEye Research Institute, Tencent’s approach is more like it has its own strategic considerations:Use external models and their own huge amount of information and users to create killer applications in the AI era, and at the same time help Tencent’s internal large models drain and iterate.
In addition to Tencent, Baidu also chose to embrace DeepSeek.
On the evening of February 16, Baidu Search stated that it would fully access the latest in-depth search functions of DeepSeek and Wenxin Model. However, on Baidu’s AI assistant Wen Xiaoyan APP, the camera problem solving function is only optimized through the DeepSeek-R1 model, which cannot be considered officially connected to the DeepSeek-R1 model.
But even so, Tencent and Baidu’s tendencies are very obvious.
In addition, among the six AI dragons, Step Star’s Yuewen APP has connected to the DeepSeek-R1 model, and the overseas version of MiniMax 01 has also launched the DeepSeek-R1 deep thinking mode.
As Tencent and Baidu’s core products WeChat, Baidu Search, and multiple AI assistants have successively connected to the DeepSeek model,The impact this has on the search market may also change the search methods and experiences of most people again in the future.
2. Independent factions: Byte Bean Bao, Kimi
In their own AI assistant applications, most companies do not have the courage to access competitors ‘AI models, except for some cases such as Tencent.
On the contrary, ByteDance, which is also a major manufacturer, even performs quite differently from Tencent.
If Tencent’s attitude towards DeepSeek is a positive embrace, then ByteDance can be called extremely conservative. Putting aside Doubao, its core AI assistant application, even other edge products are hardly connected to DeepSeek.
According to statistics, as of now, only the Volcano Engine and its products + the DeepSeek series models announced by the Flying Book official. It is normal for volcanic engine cloud platforms to be put on different large models. Therefore, strictly speaking, only one C-end product in the byte series, Feishu, has launched the DeepSeek model. Compared with the situation where more than a dozen Tencent products embrace DeepSeek, the attitudes of the two are different.
However, this attitude also seems to represent Byte’s confidence in his own big model of bean buns:Byte may believe that the big bean bag model will catch up with DeepSeek in the near future.
Before the birth of DeepSeek-R1, Doubao APP long occupied the top position in domestic AI assistant applications.
Data shows that in January 2025, the monthly activity of Doubao APP reached 78.61 million, ranking first among domestic AI applications and second in the global market, second only to ChatGPT.
However, with the explosion of DeepSeek, this pattern is being rewritten.
Little by little data shows that the daily activity trends of the three products of Doubao, DeepSeek and ChatGPT in the past 30 days can be summarized as one rise and two declines. Among them, DeepSeek’s daily activities have always been on the rise, especially on January 26 and 27. The daily activities quickly climbed from hundreds of thousands to millions. After entering February, it entered a slow upward trend.
Doubao and ChatGPT remained stable in and before early February. Since early February, daily activity has fallen precipitously and remains until now.
It can be seen that DeepSeek has now become the TOP 1 daily activity of AI application products in the global market, and after the daily activity of bean buns has dropped significantly, the gap with the lower-level ones has begun to narrow.
In addition to bean buns, the other product that has attracted the most attention after DeepSeek’s popularity is the Kimi smart assistant owned by Dark Side of the Moon.
In terms of technical capabilities alone, kimi may not be bad.
It was mentioned at the beginning of the article that the Kimi K1.5 model was released on the same day as DeepSeek-R1, with similar technical paths and praised by OpenAI.
On February 18, DeepSeek and Dark Side of the Moon collided again. They released their latest papers almost simultaneously, both themed to challenge the core attention mechanism of the Transformer architecture.
Therefore, Kimi has not connected to the DeepSeek model, but Kimi has always been criticized in terms of external reviews.
In the past, Kimi was considered by the outside world to rely too much on marketing rather than technological breakthroughs due to its aggressive launch strategy. He once paid more than 200 million yuan in a single month and was once on a hot search on Weibo.
However, despite the massive launch, Kimi still couldn’t keep up with bean buns in terms of user volume and other aspects. Now that DeepSeek has become popular, it has triggered its reflection on its re-launch strategy.
This also led to one of Kimi’s most important changes today:Adhere to the basic model SOTA and significantly reduce launch.
According to the news, after DeepSeek exploded, the Dark Side of the Moon internally reviewed that it would adhere to the basic model SOTA (State-of-the-art, the best at present). Next, it might retrain the basic model and seize application layer opportunities. One of the directions selected internally is DeepResearch, an agent product for in-depth research.
At the same time, media said that Dark Side of the Moon recently decided to significantly reduce its product launch budget, including suspending the launch of multiple Android channels and cooperation with third-party advertising platforms.
This is indeed the case:
ADX industry data shows that since February 14, the launch of Kimi smart assistants has dropped sharply. On February 15, the launch has almost bottomed out. So far, the volume of materials per day has remained at around 1500 groups, compared with the previous average of 20,000 groups. The amount of materials released dropped significantly by 92.5%.
3. Collective reflection of AI manufacturers in the DeepSeek era
DeepSeek is like a catfish, completely disrupting the global AI market in a month, triggering collective reflection from major AI manufacturers.
Liang Rupo, CEO of ByteDance, reflected that DeepSeek-R1 ‘s long-chain thinking model is not the first in the industry. In September last year, OpenAI released the long-chain thinking model and became a hot topic in the industry. Although Byte was aware of major changes in technology, it was not fast enough to follow up. If it had been the first in time, it might have a chance to achieve it earlier.
And TencentThe overall performance of DeepSeek has also reversed the outside world’s attitude towards unclear strategy and slowness in the field of AI models.
For Tencent, whether it is its own big model or DeepSeek, as long as its technical capabilities reach a certain leap-forward stage, it can quickly verify the practicality of AI technology through integration, and establish user minds in the AI era in existing core scenarios. WeChat access to DeepSeek to screen the entire network is the best proof.
At the same time, the former AI Six Little Dragons are also standing at a new crossroads.
Step Star and MiniMax chose to embrace them and connected to DeepSeek models one after another; Baichuan Intelligent continues to add AI medical tracks; Zero One Everything no longer pursues training super large models, but instead explores the industrialization of large model capabilities in commercial scenarios; On the dark side of the moon, Intelligent Spectrum AI continues to develop large models and Agent agent applications.
The grand model narrative of the past was scattered, the DAU theory was overthrown, and everyone was forced by DeepSeek to take another step forward.
For Tencent Yuanbao, accessing DeepSeek is not the same as All in DeepSeek; for Doubao and Kimi, not accessing DeepSeek does not mean a complete rejection.
Some people provide the best service status on the product side while striving for iteration time for self-developed large models; some people are unwilling to lag behind and adjust their strategies to focus on technical challenges.
This is probably the best reflection that DeepSeek, the catfish, brings to China AI manufacturers.
