Baichuan AI CEO Talks About the Price War of Large Models
The general large model has now bid farewell to simply relying on performance. Under the theory of ‘dual-drive’, its product verification has also entered a white-hot stage.
On May 22nd, Baichuan AI released the latest generation of base large model Baichuan 4, and launched its first AI assistant “Baixiaoying” after establishment.
Baichuan 4 is still a large-scale model with billions of parameters. According to Baichuan AI, compared to Baichuan 3, the new generation model’s general capabilities have increased by over 10%, with mathematical and coding abilities increasing by 14% and 9% respectively. “Baixiaoying” is positioned as a professional AI assistant, with functions including data organization, assisting in creation, multi-round searches.
Baichuan’s founder and CEO, Wang Xiaochuan, said that unlike products with the characteristics of tools in the information age, AI 2.0 turns tools into “partners,” meaning that AI can use tools like humans do, think, and have emotions. However, limited by model capabilities, related applications will gradually acquire complete abilities. In “Baixiaoying,” this pursuit is initially reflected in the application itself having abilities such as “asking questions” and “targeted search,” for example refining user questions through questioning and follow-up questions.
Both being unicorns in the domestic large-scale model field, after companies such as Minimax, Moonshot AI, 01.AI, and Stepfun successively released consumer applications, Baichuan AI’s product for ordinary users came late. In response to this, Wang Xiaochuan stated that it is not that Baichuan AI is too late but rather the industry is too early. He believes that the applications already released by the industry are just demonstrations of models and that the entire industry has not yet reached a mature state.
In fact, ‘Baixiaoying’ is just the first step in implementing Baichuan AI’s product roadmap.
Baichuan AI is a firm supporter of the theory of ‘dual-drive’ (referring to research and development and application) for large models, believing that victory can ultimately be achieved through the consumer end. Wang Xiaochuan pointed out that startups must have their own ‘super applications,’ stating, ‘If a startup company only follows OpenAI’s model by starting with a model and providing API services, it will not work in China.’
The reason for this conclusion is twofold: on one hand, he believes that in the Chinese business environment, enterprise-level businesses are ten times smaller than those on the consumer end; on the other hand, there is an irrationality in cost models – ‘You receive payment (order settlement) in RMB but spend (graphics card costs) in USD,’ as Wang Xiaochuan put it.
This statement directly addresses the recent hotly debated enterprise-side price war in the large model field. In early May, DeepSeek under the private equity giant High-Flyer Quant announced that its latest pricing for the DeepSeek-V2 API is 1 yuan for every million token input and 2 yuan for output (32K context), a price almost equivalent to one percent of GPT-4-Turbo.
This move has incited a butterfly effect in the large model field. Subsequently, Alibaba Cloud Tongyi Qwen, ByteDance DouBao, Tencent Hunyuan and other major models have followed suit with price reduction strategies for API interface services, while Baidu ERNIE Bot announced that two main models ENIRE Speed and ENIRE Lite are free.
SEE ALSO: Baidu Announces Two Main Large Models Will Be Available for Free
In this ongoing price reduction relay race among internet giants, startup companies have shown relatively low-key performance, but the spokespersons’ views are almost unanimous: startups should not blindly enter into price wars, but should instead focus on enhancing their own model performance.
Regarding his views on price wars, Wang Xiaochuan believes that “everyone is really optimistic about the prospects of this era and unwilling to miss any opportunities, which indirectly reflects everyone’s sufficient yearning for AI capabilities in this era.” Furthermore, he judges that cloud providers may seize the opportunity of large models and even potentially break free from the industry’s previous dilemma of unclear profit models.
According to his understanding, the essence of this round of price reduction by major companies is that cloud providers are entering a new battlefield. The price reduction is not only within the range of these major companies, but also limited to actions taken by cloud providers. If it is targeted at enterprise services, this war will eventually turn into selling a complete set of cloud services rather than just the model itself. This story is very similar to what happened with the “AI Four Dragons” (usually referring to SenseTime, Megvii, Yitu and CloudWalk).
Wang Xiaochuan claims that he does not feel any anxiety due to the ongoing dispute in front of him, but he advises startups not to get involved. In his view, this is not equivalent to burning money like Didi and Meituan did during their time; it cannot change the production relationship based on supply-demand bilateral networks. At the same time, “do not make such a business model (referring to enterprise-side models represented by open API interfaces) your focal point; this logic does not drive a startup company with dual wheels.
Compared to the fierce competition in the enterprise market, although there is currently no price war in the consumer market, a marketing battle involving start-ups buying traffic and expanding their presence has emerged.
SEE ALSO: Baichuan AI Releases Large-scale Model Baichuan 3 with Parameters Exceeding One Trillion
In response to this, Wang Xiaochuan still believes that this is not a healthy behavior and may even be just a means to accelerate the financing process. Large companies have different paths to choose from in terms of product and marketing coordination – some focus on developing models first while others prioritize applications. However, finding a balance between models and applications is a top strategic consideration for every company.
As for the sign of the arrival of the “super app” era, Wang Xiaochuan’s definition is to increase the current daily active users by two orders of magnitude. If we take 1 million as a benchmark, then a “super app” will be a product with daily active users in the hundreds of millions.
Regarding the business model of “super apps,” he believes it is too early to discuss now and bluntly states that existing API revenue and model revenue (such as membership subscriptions) are not the most attractive business models on the market.
“If you have a super app that can solve user pain points, from an income perspective, I am optimistic without any doubt,” he said.