GuShiio.com Article source: Fixed Focus One
Image source: Generated by AI
AI seems to have become the “life-saving straw” for big factories.
Whether it is the highlight data in the financial report or the good information every now and then, it is inseparable from AI.
For example, in Baidu’s mixed financial report in 2024, the highlight moments are basically given by AI:
The average daily call volume of Wenxin Model continues to grow rapidly, increasing 33 times a year to 1.65 billion yuan. Baidu Library has more than 40 million paying users, ranking second in the world and first in China.
Ali also relied on AI to make three consecutive hits at the beginning of the New Year:
First, influenced by DeepSeek, Qwen, which is also a big open source model, attracted attention; then the latest model Qwen2.5-Max was released, which was evaluated as exceeding DeepSeek V3 in performance; then it was announced that it had reached a cooperation with Apple on the AI business, and its share price soared.
However, in the nearly 40 days since DeepSeek came out of the industry, the AI of the big factory has suffered more anxiety than the gain. After all, each family has invested a lot of manpower, material resources, and financial resources. In the end, the blockbuster product was the product made by a start-up team.In the past two days, DeepSeek also disclosed explosive news for the first time-its cost-profit margin is as high as 545%(theoretical income), and its profit can theoretically reach 3.46 million yuan per day.
Under various shocks, major manufacturers have changed their routes one after another, joining if they cannot win, and announcing access to DeepSeek. At the same time, they have shifted their large models from closed-source to open source, and even cut off a commercialization path and free C-end products.
However, can this wave of operations really cure the AI anxiety disorder of big factories?
How is the big factory AI?
Before the emergence of DeepSeek, the route for large factories to make AI was to focus on high-level efforts, focus on investment, and make products based on their own advantages.
The big model is regarded as the infrastructure of the AI industry. Major Internet companies (Baidu, Tencent, Ali, Byte, Fast Hand, etc.), consumer electronics manufacturers (represented by Huawei), and intelligent voice manufacturers (iFlytek, etc.) have all launched self-developed big models. Compared with startups such as “AI Six Tigers”, Dachang has the advantage of having stronger capital and talent reserves.
Judging from the overall technical iteration speed of the AI industry and the public information of each family, there is no fundamental difference in the underlying technology of the big factory model, but there are differences in entry time, model positioning, and market strategy. The specific differences are as follows:
These three differences represent to a certain extent the early attitude and positioning of major factories towards AI.
For example, when a large model is released,”early” means that the big factory has an early layout and technology accumulation in relevant technology fields, and has a rapid response. However, the risk is that the technology is not yet fully mature, and the investment in technology research and development and marketing costs are relatively higher.
Judging from the above table, Huawei was the earliest, but it should be noted that although its underlying layer is also based on the Transformer architecture, it is completely different from ChatGPT-style dialogue and belongs to the AI model in the “industry-specific” direction (ChatGPT-style is universal intelligence). If it focuses on the general intelligence model, it will be Baidu’s earliest action. In March 2023, it will launch an invitation for testing of the Wenxinyiyan model (not fully open).
However, launch time is not a core factor in measuring the quality of a model.
The business layout of a large factory determines the application direction of large models, and also creates the positioning of different large models. Technically, they come from the training data of each family.
Baidu Wenxin’s words mainly rely on Internet text data; Ali Tongyi Thousand Questions is multimodal data such as text, pictures, and audio; Tencent Hunyuan is data on Social networks and user behavior; About 50%-60% of the bytes come from byte’s own business (Douyin, Today’s Headline) data; Huawei Pangu Model uses various data including industry, meteorology, text pictures, and images.
This also makes each model have different advantageous scenarios. For example, Wenxinyiyan dominates long text processing and multilingual dialogue; mixed yuan is better in social scenarios; bean buns are more advanced in generating content and accurate recommendations; Tongyi Qianwen responds faster in e-commerce recommendation scenarios; Pangu has excellent execution speed and generalization ability, and can efficiently respond to large-scale tasks.
It is not difficult to find that the advantageous areas of each big model all have their core business shadow.
last lookmarket strategyTo a certain extent, what reflects is that major manufacturers ‘judgments on their own capabilities and industry trends are divided into two parts: whether opening and closing sources and TOC products are free.
Byte, Fast Hand, iFlytek and Huawei are still insisting on closed source, while Baidu, Tencent and Ali choose most of them open source. On the TO C application, Baidu, Tencent, and Ali have chosen free routes, and Byte, Fast Hand, and iFlytek most provide a limited number of free quotas.
The sweetness of open source has been enjoyed by Ali. The latest list of open source models released by the open source AI platform Hugging Face shows that the top ten open source models are all derivative models based on Alitong Yiqian Wen.
Among TO C products, bean buns that insist on free have seen the strongest rise in the year. According to the AI product list, in January 2025, bean buns ranked first among the tens of millions of monthly clubs in China, with 78.61 million, far exceeding the applications of other major manufacturers.
However, what everyone is more curious about is the ranking of the overall capabilities of large factories and large models. According to analysis by many practitioners, the current top large models of big factories are mainly closed-source. When the information is not completely transparent, it is not easy to judge the capabilities of each company.
Frost Sullivan pointed out in the “2024 China Large Model Capability Evaluation” report that large models such as Baidu Wenxinyiyan, Tencent Hunyuan, and Alitong Yiqian are all in the first echelon, believing that they are relatively comprehensive in technical capabilities and have a relatively large number of users. However, no clear judgment was given on which one had the better overall ability.
I software engineer Qin Xiang said that there are differences in technical architecture and training data among each family. For example, from the perspective of technical architecture, model size and parameter quantity are important indicators to measure the complexity and capabilities of large models. Generally speaking, the larger the scale and the more parameters, the stronger the learning ability and expression ability of the model will be. For example, DeepSeek-R1 is known as the giant in terms of parameters, and its 671 billion parameters create a huge knowledge pool.
He said that judging from this dimension, among the large models in large factories, large models with in-depth reasoning capabilities, such as Wenxinyiyan, are more capable among large factories. However, if you look at the capabilities of vertical domains, Wenxin’s words are not as good as Tongyi Qianwen. After all, the latter has developed and launched eight vertical domain models based on itself.
Anyway,The advantages of each large model are different, and it is difficult for one to crush others in all dimensions.
DeepSeek has been out of the circle for 40 days, and four major changes in a large factory
The emergence of DeepSeek has prompted major manufacturers to re-examine their AI strategic layout. Combined with the latest developments and practitioners ‘opinions, there have been four specific changes.
The first is from closed source to open source, which is also the most significant change.
More than one practitioner pointed out that DeepSeek’s popularity is inseparable from open source.
There has been constant discussion at home and abroad on opening and closing sources for large models. Robin Li, chairman of Baidu, was once a loyal supporter of closed sources and believed that whether it maintains its technological leadership or business model, closed sources are stronger than open source.
Qin Xiang analyzed from a technical perspective that open source means that the core code is open, and competitors can quickly replicate the technical path. In the early days, big manufacturers chose to close source mainly to protect intellectual property rights and commercial barriers (for example, OpenAI did not open source GPT-3 in the early days).
But he found that led by DeepSeek, big manufacturers have changed direction and are more inclined to achieve long-term benefits through ecological bindings (such as Tencent’s mixed open source video model to attract developers to use its cloud services) rather than relying solely on technical confidentiality as before.
Now Baidu has announced that the Wenxin Model 4.5 series will be fully open source by the end of June 2025. As of now, most of the models of Baidu, Alibaba, and Tencent have been open source or announced to be open source.
Second, the business focus has shifted from TO B to “two-line parallel”.
Qin Xiang explained that there are three main ways to monetize large models: value-added services, data monetization, and compliance services. Among them, value-added services account for the largest proportion, relying on enterprise-level customization and API call revenue. He revealed that the annual fee for Baidu Wenxinyiyan Enterprise Edition exceeds 10 million yuan, and Alibaba Yuntong Yiqian provides customized customer service systems for government and enterprise customers, with a single project contract value reaching hundreds of millions of yuan.
In other words, the current profits of major manufacturers still rely mainly on B-terminals, but recently many major manufacturers have begun to pay attention to the promotion of TO C applications and changed to “two-line parallel” to TO B and TO C.
Photo source/ Pexels
For example, Tencent has increased its promotion of Yuanbao. On the one hand, it has connected it to WeChat Jiugongge, which has a strong traffic portal. On the other hand, it advertises through multiple channels. In addition to promoting it in Tencent’s ecological industry, Douyin, Station B, Zhihu also made a large number of releases.
According to App Growing data, among the top 20 AI tools with advertising intensity in February, major AI products were on the list (Huawei has no TOC products that are not among them). Among them, Tencent Yuanbao spent the most money. In February this year, its investment amount accounted for 46% of the total amount, which was almost up to the total of the past nine months and exceeded the number of bytes of bean buns.
In addition, Ali is also recruiting talents related to TO C business on a large scale.
Practitioners believe that it may be that DeepSeek’s low open source +API prices have brought greater pressure on the TO B business of the big factory, and then want to find more commercial outlets on TO C.
The third change in direction is that TO C applications will change from charging to free.
DeepSeek is easy to use and free. After its popularity, Wenxinyiyan from domestic Baidu and GPT-5, which will appear on foreign OpenAI, announced that it will be available to users for free.
“The purpose is to attract more users and increase market share.” Qin Xiang said that more user feedback can further optimize model performance, thereby improving B-side service capabilities and charging higher enterprise customized model fees.
The fourth transformation is to fight price wars from heavy investment to cost reduction.
In the “Hundred Models War” in the past few years, domestic and foreign AI model companies have spent billions or even tens of billions of dollars, but DeepSeek has trained the DeepSeek R1 model with the capabilities of OpenAI o1 at a GPU cost of only US$5.576 million, which has made big manufacturers start to reflect.
Qin Xiangneng clearly feels that since last year, the competition for large models has shifted from “technology first” to “cost + ecology”. For example, in January last year, the price of the API released by the Bean Bag 1.5Pro dropped sharply. In December, Byte reduced the price of visual models to 85%, pushing the industry into the “era of centimes.”
In February this year, two old Baidu people were also fighting from afar over the price of large models. Shen Dian, president of Baidu Intelligent Cloud Business Group, pointed out at the all-staff meeting of Baidu Intelligent Cloud Business Group (ACG) that there was a “malicious price war” in the domestic large model industry. Then Tan, president of Byte Volcano Engine, responded in a circle of friends, pointing out that price cuts are the inevitable result of technological progress.
DeepSeek was not idle either. It had just announced the end of the API promotion period. On February 26, it announced a “limited time price reduction”. From 00:30 to 08:30 every day, DeepSeek V3 was reduced to 50% of the original price, and DeepSeek R1 was as low as 25%, with a maximum drop of 75%.
The pressure on big factories is even greater.
Free and open source, can big factories win back their home games?
According to practitioners, among the four major changes, the ones that currently have the greatest impact on large manufacturers are open source and free.
Let’s first look at open source.
Liu Cong, an expert in the field of large models, pointed out that before DeepSeek showed up, whether foreign OpenAI or domestic manufacturers, either chose to close all sources or open source some large models (not the best versions). DeepSeek used its most powerful reasoning model DeepSeek-R1 also chose open source, which is what practitioners are very excited about.
ButOpen source also faces some lost revenue and technical risks.
Weiliang, a doctor in artificial intelligence, said that open/closed source represents two business models and development ideas: indirect/direct monetization. A typical open source representative of domestic manufacturers is the Alitongyi Thousand Questions Model, which further promotes commercial cooperation by adapting manufacturers. This move is a choice based on its own ecology.
However, many major manufacturers initially positioned themselves as technology-oriented and regarded them as productivity. For example, OpenAI, Baidu, Huawei, and iFlytek of Science and Technology. Large model subscriptions are a very important source of income, and choosing open source will definitely affect profits.
Open source also faces risks of malicious attacks and community maintenance. For example, under code disclosure, malicious attackers can analyze the code to find vulnerabilities to attack systems using these models.
Subsequent community maintenance is also a problem. Qin Xiang said that open source needs to continue to invest resources to maintain the developer community (such as providing documentation, technical support, version updates), otherwise it may lead to the decentralization of the technical ecosystem. He explained that if developers modify their own code and generate multiple branches (such as Linux branches Ubuntu and CentOS), it will make it more difficult to unify technical standards and lead to “technical fragmentation.”
Photo source/ Pexels
Some practitioners bluntly said,Even if big manufacturers open source, their appeal to them is limited.
The purpose of open source is to attract technology developers and partner companies and let everyone use its large model for technical iteration and application development. However, Dr. Weiliang believes that “currently, open source companies are suspected of advertising.”
“What open source can see is the reasoning methods and parameter weights of large models, but everyone has not liberalized the more important data screening methods and model training skills, which also makes it difficult for ordinary developers to do technical iteration.” He said.
It is worth noting that open source does not mean all free. Users must also fulfill the open source agreement of the large model provider, which includes a “payment clause”.
For example, Dr. Weiliang will use the Alitongyi Thousand Questions Model to do some AI applications. After using Thousand Questions to run the technology, if you want to further customize fine-tuning and adaptation, you need to contact the staff. He also revealed that open source agreements will also have restrictions such as company size, such as payment when the number of employees reaches a certain number.
Let’s take a look at the impact of free.
The purpose of adopting a free strategy is to quickly occupy the C-end market. For example, the prominent representative is the bean buns that have always been free to users. QuestMobile data shows that as of February 9, 2025, the average number of active users on the week of bean buns (based on February 3-February 9 is the week, and the average number of active users per day is calculated). The number of active users is 18.45 million, second only to DeepSeek and higher than Kimi, Wen Xiaoyan, Tongyi, and Yuanbao.
However, practitioners are not sure how much free means. This is not only because users have low loyalty to tools such as chatbot, but also because domestic users are not aware of payment.
“Even for AI-generated video tools that require payment, most domestic applications rely on providing free points to attract users to use them.” One practitioner said that he felt that bean buns could be found among a number of similar general-purpose AI products. In addition to being free, it was also inseparable from byte’s powerful marketing.
Qin Xiang believes that DeepSeek’s catfish effect has forced large factories to shift from technical competition to a comprehensive competition between cost and ecology.. Even if these measures will reduce their own profits in the short term, they have to take them.
The catfish effect triggered by DeepSeek is not over yet.
