Image source: generated by Unbounded AI

Kunlun Wanwei is the most special company among the large model card tables that have been shuffled repeatedly.

Compared with a number of large companies with a market value of hundreds of billions, such as Alibaba and Baidu, Kunlun Wanwei can only be regarded as a small and medium-sized enterprise. But in the large-scale model track, which burns both resources and talent, Kunlun Wanwei not only got a share of the pie in the AI ​​field first. Two years later, Kunlun Wanwei, which was questioned for "following the trend", has outlasted public suspicion.

On January 6 this year, Kunlun Wanwei officially launched the o1 and 4o versions of the "Tiangong Model 4.0" based on the OpenAI o1 model. Officials stated that this is the first Chinese logical reasoning model in China. The o1 model with capabilities, the latter benchmarks against GPT-4o, achieving a breakthrough in end-to-end voice.

O1 version of "Tiangong Model 4.0" Source: Kunlun Technology

Why can Kunlun Technology be put on the table? The conclusion given by founder Zhou Yahui is timing.

“As you can imagine, if we hadn’t done the (large model) very early and persisted in doing it, we would definitely not have any position now.” Talking about Kunlun Wanwei’s advantages in the AI ​​field, Zhou Yahui gave gave this answer.

But a year later, Kunlun Wanwei’s accumulated time advantage on the large model track gradually became less obvious. After the limited dividends gained during the hot period passed, the company began to turn to a more possible field - AI applications.

To make a Super App, technical advantages and product thinking are indispensable. When Scaling Law slowed down and the capacity iteration of large models slowed down, Kunlun Wanwei, which started out as a product, began to try to leverage its product and operational advantages in the Internet era, hoping to survive on the large model table longer.

Will it succeed?

Two years to get a head start on AI

Given one and a half years ago How did Zhou Yahui, who scored 40 points for Byte's AI strategy, lay out Kunlun Wanwei's AI strategy?

The most important thing is to be fast enough.

“This wave of large models is still in the very early stages,” said Fang Han, chairman and CEO of Kunlun Technology, in October last year. “Kunlun Technology is trying more in multiple directions and trying quickly. Fast fail (fast iteration), quick change of direction”

In more than two years of “All in AGI and AIGC”, Kunlun Wanwei laid out its plans from the large base model to the construction of the application layer ecosystem. The speed largely determines this company’s position in the AI ​​industry.

In 2023, Kunlun Wanwei took the lead in studying large models. It was one of the first players to join the "Battle of 100 Models", and thus reaped the dividends of market value growth: 4 In September, the benchmark GPT-3.5 model was launched.After the "Tiangong" model was unveiled, Kunlun Wanwei's share price soared 230%, and its market value also soared, reaching a peak of 79.3 billion yuan.

Changes in the market value of Kunlun Technology in the past three years

When it comes to AI applications, Kunlun Technology has also continued the idea of ​​​​"quickly trying". An employee close to Kunlun Wanwei commented to Guangcone Intelligence, "Boss Zhou has a unique vision and quick thinking, so Tiangong AI started early."

Kunlun Wanwei is currently the AI ​​application with the largest number of users As for "Tiangong AI", it is currently one of the few AI products in China that has entered the "five million club" with monthly active users. As a comprehensive assistant application, Kunlun Wanwei has bet on the AI ​​search track early and became the first company in China to launch an AI search function. An employee of Kunlun Wanwei told Guangcone Intelligence that at that time, Tiangong AI was also the first application of its kind to be launched.

During the blank period of AI search, Kunlun Wanwei also benefited from the early market dividends and accumulated a number of users by taking the lead in developing products and investing in them.

According to statistics, from January to October 2024, the top 5 cumulative downloads of China’s AI applications (APPs) include Doubao, Wen Xiaoyan, Kimi, Hoshino and Tiangong AI. Among them, Tiangong AI has accumulated 11.7 million downloads in 10 months.

In the statistics of new AI native applications, Tiangong AI ranks fifth. Source: AI product list

However, the advantage of early start can help Kunlun Wanwei gain a foothold first. But there is no guarantee that it will always stand on the card table.

Following the layout idea of ​​large infrastructure models-AI applications, Kunlun Wanwei is still optimizing the large model capabilities of the base, catching up with the pace of large model teams at the head, and laying a good foundation for the AI ​​application matrix. Technical basis.

In November, Kunlun Wanwei launched Skywork o1, ​​a large model that benchmarks OpenAI o1, and released an open source model based on the open source Llama 3.1 8B and two dedicated versions, Skywork o1 Lite and Skywork o1 Preview.

After adding Q* algorithm and reinforcement learning capabilities, Kunlun Wanwei’s open source model version has significantly improved capabilities in terms of mathematics, coding and other abilities. In the field of mathematics, it has made the base model Llama-3.1-8B The capability reaches SOTA, surpassing Ali's Qwen-2.5-7B instruct.

In terms of multi-modal capabilities, Kunlun Wanwei is also following the "small steps and fast pace".

This month, Kunlun Wanwei’s multi-modal large model Skywork 4o, which benchmarks GPT-4o, is officially launched. While it can respond efficiently, it supports real-time interruption, emotional understanding, personalized memory, and support The ability to customize sounds and more.

However, in terms of the market position of basic large models, Kunlun Wanwei’s once-leading position is becoming less obvious. In terms of multi-modality, many domestic companies have already grabbed theGet the upper hand. Taking end-to-end voice as an example, Zhipu AI, iFlytek, Xihu Xinchen, etc. have released their own end-to-end voice models and products as early as September. Kunlun Wanwei is nearly 3 months late.

Kunlun Wanwei’s current situation in GPT-4o products is to some extent a reflection of the company’s current situation: in the large model industry, gods are fighting, and the industry is accelerating. Under such circumstances, Kunlun Wanwei's speed gradually lost its advantage.

Basic large models are slowing down and they want to use AI products to overtake others

Whether in the field of large models or at the AI ​​application level, the funds of major manufacturers , The traffic advantage is the moat for them to increase their investment in AI, and it is also an insurmountable barrier for small and medium-sized companies. Kunlun Wanwei is no exception to this point.

In the field of large models, Kunlun Wanwei has begun to fall into the same situation as its own position as a "medium-sized company": it is difficult to compete with large manufacturers and new entrepreneurs, but it still has a place in self-developed models. .

From the perspective of multi-modal model updates with comprehensive capabilities, the latest version of Kunlun Wanwei is the Tiangong Model 3.0 released in April this year. According to official information, the model parameters are as high as 400 billion, which has been achieved It surpassed Grok and became the largest open source MoE model in the world at that time.

However, with the continuous follow-up of domestic large model companies, the open source large model capabilities in the multi-modal field are updated and iterated extremely quickly. Just two days ago, Deepseek's V3 large model was officially launched. On most benchmarks, its capabilities have surpassed the top closed-source model GPT-4o.

With the slowdown of Scaling Law, the effect of pre-training on large base models is limited. Relying on pre-training alone, the capabilities of large models will no longer increase rapidly.

“It took us less than a year to go from belief to doubting Scaling Law.” After making the decision to split the pre-training team, Kai-fu Lee commented on the changes he saw in the development of large models. When it comes to training very large models, his judgment is that the cost and price will be very high, and it is definitely not something a start-up company can do.

In addition, the cost-effectiveness brought about by the increase in the technical capabilities of large models has been compromised. As Fang Han judged at an AI conference, SOTA’s dividend period is becoming increasingly short.

Thus, another path that Kunlun Wanwei has paved early, its upper limit is becoming wider.

“The innovation of AI products in product models and business models lags far behind technological innovation.” Fang Han said, “However, for C-end users, the product and business Model innovation is more important.”

In the past two years, Kunlun Wanwei has maintained a rapid iteration rhythm in the five major fields of AI short dramas, AI social networking, AI search, AI music, and AI games. Corresponding products were released.

In Kunlun Wanwei’s plan, the above five major platforms can be summarized into two business lines——AI search and AI UGC platform. Fang Han believes that AI search will become a super APP that improves the work efficiency of C-end users, and the AI ​​UGC platform will become a space for users to express themselves. These generated assets will also be the key to improving the connection between users and applications.

Whether it is a search or a content platform, for Kunlun Wanwei, relying on its previous acquisition of Opera, the third largest overseas browser, coupled with its early overseas entry and a batch of overseas social product matrix, it We have a rich reserve of products and operational experience in related applications.

The advantages are also shown in product data. It is worth noting that in the two fields of AI search and AI social applications, Kunlun Wanwei’s accumulated product capabilities and user base have the opportunity to equal the big manufacturers and the “Six Little Tigers”. Tiangong AI currently has more than 5 million monthly active users; as of October last year, its AI social product Linky had more than 3.78 million overseas monthly active users, ranking sixth on the list of AI overseas products.

But in the field of AI applications, the difficulties Kunlun Technology faces in some subdivided tracks are no different from the development situation of its large model companies.

As far as search is concerned, although Kunlun has the advantage of building its own index database, this alone has surpassed most AI search manufacturers on the market. However, major manufacturers that also have browser product accumulation and more data and traffic resources can easily complete the early user accumulation.

Take Baidu's "AI Search" as an example. In December, the number of visits to the product's web page increased by more than 90% month-on-month, with more than 11 million visits in a single month, twice the number of visits to Kunlun Tiangong AI. above.

Image source: AI product list

This month, Baidu’s “AI Search” was officially launched. It also embedded the function on the homepage of Baidu Search. Traffic portal, more users will be imported into Baidu’s new AI search products. Even in mobile hardware terminals, search players such as Baidu and 360 tightly control market share.

When big-name players end up doing the same business, Kunlun Worldwide doesn’t have much time left.

Kunlun Wanwei grasps the straw of AI

Looking at Kunlun Wanwei’s business changes from its establishment to the present, it seems that the company can always smell the wind and seize it at critical moments.

It started with online games and went public, then acquired browsers and decided to go overseas. Kunlun Wanwei relied on the game business to build a foundation. When the revenue of this part declined, it jumped into the "blue ocean period" of going overseas. The ship leading to overseas markets - In 2010, Kunlun Wanwei turned its attention to the world. Through the acquisition of a series of applications such as Grindr and Opera and the independently developed product matrix, it established a foothold in overseas markets early.

Nowadays, AI has become Kunlun Wanwei’s third card.

“As a medium-sized Internet company, our only dream is to become a giant Internet company, so we have been looking for the second curve.” When talking about why we want to build AI, Fang Han gave this answer.

If you succeed, you will use AI to get on the table and move towards becoming a large company. Failure, Kunlun Wanwei still maintains the mid-range level, once again looking for the next new outlet worth betting on.

This is not the first time that Kunlun Wanwei, which has been questioned for its ability to adapt to the wind, is looking for a new growth curve. However, compared to the metaverse, new energy and other businesses that it has left behind that are too far away or not close to its own development, AI is undoubtedly the best solution that Kunlun Wanwei can grasp at the moment.

That’s why Kunlun Wanwei dared to shout the slogan “All in AGI and AIGC” and spent a lot of money on AI. Kunlun Wanwei pointed out in its third quarter financial report that in order to speed up the application of AI technology, the company's research and development expenses in the first three quarters increased to 1.144 billion yuan, a year-on-year increase of 84.47%. According to Fang Han, in order to reserve sufficient computing power, Kunlun Wanwei spent 800 million on purchasing chips in 2023 alone.

For Fang Han, these losses are a "necessary path" in his eyes. As technological development gradually improves, Kunlun Wanwei's losses will further narrow.

“We are still in the early stages. To adapt to the market, we must make relatively large investments. I believe that these investments will definitely obtain sufficient returns in terms of market share in the future.” Fang Han said, "All inference and training costs are declining sharply, and the company's consumption is also declining. The losses should not continue."

Many securities companies gave Kunlun Wanwei "buy" and "hold" rating. Huatai Securities pointed out in a research report that it is optimistic about Kunlun Wanwei's competitive advantages in AI large models and applications. The application side is expected to continue to be implemented, maintaining a "buy" rating.

In 2025, more people in the industry are betting that this will be the year of the explosion of AI applications - when the cost of AI calls is infinitely low and manufacturers can be responsible for their own profits and losses, AI applications have the potential to develop Basic characteristics of the audience.

But from the first day of moving towards AGI, many domestic companies, including Kunlun Wanwei, have been concerned about commercialization issues. If you can quickly trial and error from the huge AI product line and find a Super App that can support the company, all the losses in front of you will be worth it.

On the contrary, Kunlun Wanwei can only give up this piece of fat that once increased the company's market value. As Fang Han judged, for Kunlun Wanwei, the criterion for abandoning AIGC is not how much money is burned, but if it is not in the first group, there is no need to continue.

Kunlun Wanwei, which has laid the foundation for multi-modal large models, has already paid a lot of "sunk costs". To this end, in addition to sticking to the existing AI search and AI UGC, it will continue to launch a series of AI application products to realize these costs.

Fang Han’s attitude towards AIGC and AGI is also the situation that all domestic AI companies must face:

“AGI is definitely our (Kunlun Wanwei) goal. But I have to Just live on first,AIGC Let’s survive first. ”