iFlytek Secures Hundreds of Millions of Yuan in Contracts, Huawei Collaborates with Stock Exchanges. Are the Big Model Players Quietly Making a Fortune?
On November 29th, iFlytek Medical won the bid for the software service of the Hefei National Artificial Intelligence Application Pilot Base (in the direction of primary health services in the medical field) project, with an order amount reaching 430 million yuan.
It is reported that this project involves the procurement of 6 large models. Among them, 1 is a general large model, and 5 are custom - developed large models in the medical field. Specifically, they are the general cognitive large model (finished software), the medical cognitive large model (custom - developed), the medical voice large model (custom - developed), the medical imaging large model (custom - developed), the medical image - text recognition large model (custom - developed), and the medical multi - modal large model (custom - developed). The tender price limits are 16 million yuan, 21.555 million yuan, 4.8 million yuan, 56.22 million yuan, 7.68 million yuan, and 23.775 million yuan respectively.
According to the tender announcement, in addition to procuring the general large - model base and 5 domestic, independently controllable, large - scale medical vertical large models with hundreds of billions of parameters, such as medical cognitive, voice, imaging, image - text recognition, and multi - modal models, the project will also build a high - quality medical data resource platform. It will collect, aggregate, and manage no less than 5PB of medical and health behavior data, and build a medical data resource center with 34 high - quality medical knowledge bases for medical large - model enhancement, 42 high - quality medical multi - modal data sets for medical large - model construction, and 11 data sets for primary medical disease research and transformation. The medical large model, combined with the data resource platform, can comprehensively accelerate the implementation of AI + medical service applications.
Previously, on September 29th, iFlytek won the bid for the "Wucheng Digital Intelligence Future" new infrastructure construction project in Jinhua, with an order amount reaching 380 million yuan. This project covers ten industry AI implementation scenarios, including AI + education, AI + public security, AI + human resources and social security, AI + investment promotion, AI + industry, AI + medical, AI + water affairs, AI + management, AI + office, and AI + court.
After a detailed statistical analysis of the bid - price documents, the software part of this project (accounting for 41%) exceeds the hardware part (accounting for 38%). This has also been regarded by some media as "an important signal that AI large models have started to 'make money'."
Applications of vertical large models are becoming increasingly "valuable." Coincidentally, one day before iFlytek won the latest large order, Huawei and the Shenzhen Stock Exchange jointly released a large model for the securities industry regulations based on Huawei's Pangu Reasoner 38B:
By integrating the industry regulations knowledge system of the Shenzhen Stock Exchange (more than 100,000 regulatory clause label data) and combining Huawei's self - developed fine - tuning technology and scenario - based training capabilities, the large model can achieve a question - answering accuracy rate of over 90% and has features such as scalability and real - time performance.
Winning Hundred - Million - Yuan Large Orders, Vertical Large Models Take the Lead
Compared with vertical large models, the public is undoubtedly more familiar with general large models such as Doubao, Qianwen, Yuanbao, and DeepSeek, which are commonly used. However, at present, vertical large models are more likely to make money than general large models.
The core reason is that vertical large models have clear goals, which are to improve the work efficiency of various industries in dealing with specific practical problems. They can help industry users solve the long - standing problem of "cost reduction and efficiency improvement." This business path is not only clear and direct but also can be monetized more quickly and easily. They can be regarded as a type of "professionals with specialized skills."
General large models have lofty aspirations, aiming to build an "anthropomorphic" communication role that knows everything and can do everything. While actively embracing the general public, they do not explicitly focus on making money. They can be regarded as a type of "talents with extensive knowledge."
(Image source: iFlytek Medical)
From the perspective of content output, vertical large models are better at solving practical problems and handling specific tasks, acting as human assistants; general large models are better at inspiring thinking and generating creativity, acting as human companions.
Assistants are for making money, and companions are for companionship. From positioning to function, vertical large models are more likely to make money than general large models.
However, without the base or the "parent body" provided by general large models, vertical large models would become "water without a source and a tree without roots." Therefore, we can see that the medical project just tendered in Hefei is a combination of 1 general cognitive large model + 5 medical application large models (with different focuses).
Deeply Engaged in Vertical Large Models, Many Enterprises Are Quietly Making Money
Not only technology giants such as iFlytek, Huawei, Alibaba, ByteDance, Tencent, and Baidu, but also more players from different sectors have flocked into the domestic vertical large - model market, including large enterprises in vertical industries and many new large - model startups, and they are all quietly making money. In terms of industry categories, customers in enterprise or public sectors such as finance, medical, education, and government affairs have become the first batch of "big spenders" in the vertical large - model market.
In the financial field, due to the large number of segmented application scenarios and the ability to achieve quick monetization, not only technology giants (such as Ant Group and Baidu) are actively involved in the R & D and promotion of financial large - model products, but also many securities, banking, and insurance enterprises are participating in the name of "tech - finance."
In the segmented application scenarios of financial large models, intelligent investment research and advisory, intelligent risk control and compliance, intelligent marketing and customer service, intelligent trading and operation, etc. have become hot areas. It is worth mentioning that Magic Square Quantization, the parent company of DeepSeek, is a private equity institution focusing on quantitative investment and AI technology R & D. In the A - share market, the current amount of quantitative trading has reached a considerable scale.
Magic Square AI's deep - learning training platform "Firefly II" (Source: Magic Square Quantization)
For current large - model startups, vertical large models are also a better way to "make a living" than general large models. After all, one must survive first before pursuing lofty goals.
Among the "Six Little Tigers of AI Large Models" enterprises, Zhipu AI, Lingyi Wanwu, and Jieyue Xingchen have also been focusing on the vertical large - model field this year, or accelerating commercial monetization through the B - end application market.
Among them, Zhipu AI not only maintains its presence in the C - end application market but also promotes the implementation of large models in multiple B - end vertical industries such as finance, medical, education, and government affairs. It claims to have tens of thousands of partner enterprises. It is worth mentioning that since this year, this Beijing - based AI unicorn enterprise has also received strategic investments from state - owned assets or industrial funds in many places across the country, such as Hangzhou, Zhuhai, Chengdu, Beijing, and Shanghai. It may have won the representative position of large models for government services among AI startup teams.
Lingyi Wanwu, founded by Kai - fu Li, announced this year that it will no longer be involved in ultra - large basic models with over a trillion parameters. Instead, it focuses on bridging the "last mile" from the base model to vertical scenarios and has launched the Wanzhi Enterprise Large - Model One - Stop Platform.
Jieyue Xingchen is currently one of the enterprises with a relatively lower valuation among the "Six Little Tigers of AI Large Models." This year, Jieyue Xingchen began to cut its C - end product line and focus on the R & D of AI agents and large models. Some media interpret this as a shift in the company's development focus from covering C - end product applications to targeting the developer community and the B - end industry market.
Verticalization of General Large Models, Accelerated Iteration of Vertical Large Models
Since this year, with the increasingly aggressive general large - model products and the continuously popular AI agent technology applications targeting the C - end user market (note: there are also a large number of AI agent technology application products in the B - end large - model market), the anxiety of vertical large - model enterprises and industries has also been increasing. Many people have begun to pessimistically believe that vertical large models will be completely replaced by general large models in the near future.
This view is not without reason, but the replacement process may not be that fast. As shown in the large medical order project in Hefei that iFlytek just won (1 general cognitive large model + 5 medical application large models):
In industry - or enterprise - level AI platform applications, general large models can be responsible for understanding, cognition, planning, and interaction, while vertical large models can be responsible for performing specific sub - tasks, thus forming a collaborative network system of "one general brain commanding multiple professional 'cerebellums'."
Therefore, in the foreseeable future, both general large models and vertical large models will have a wide range of applications. In the industry - and enterprise - level AI application market, in addition to vertical large models highly adapted to various industries, the base and public - attribute capabilities of general large models are also needed.
Meanwhile, players focusing on the vertical large - model field, especially startups, are also accelerating the technological and application iteration process of vertical large models.
In the medical large - model field, in addition to large enterprises like iFlytek (which has even established its subsidiary iFlytek Medical), there are also many startup enterprises. As an early - stage startup in the vertical large - model field and also one of the "Six Little Tigers of AI Large Models," Baichuan Intelligence, founded by Xiaochuan Wang in 2023, after two years of exploration and one year of adjustment, further established its company development strategy of focusing on AI large models and their applications in the medical field this year.
In August this year, Baichuan Intelligence launched a new medical large model, Baichuan - M2. It has only 32B parameters, but it outperforms open - source/closed - source models several times its size in various benchmark tests. Designed for real - world medical reasoning tasks, Baichuan - M2 supports single - card deployment on RTX4090, which means that even small and medium - sized medical institutions have the conditions for private deployment.
Ranking of large - model evaluation test sets in the medical and health field in August this year (Source: OpenAI)
Xiaochuan Wang said that the performance of this large model exceeds that of two open - source models previously released by OpenAI. In the closed - source field, its ability is second only to GPT - 5. Xiaochuan Wang does not hide his new goal: To surpass general models in this vertical medical field.
Continuing to strive, in October this year, Baichuan released Baichuan - M2 Plus. Through the first - created "six - source evidence - based reasoning" paradigm, it focuses on improving the professionalism and credibility of medical large - model responses.
Specifically, in the knowledge collection stage, Baichuan - M2 Plus will actively block non - professional information sources on the Internet and only use medical evidence from authoritative sources to ensure that responses are based on "high - level evidence." At the same time, it divides medical evidence into six levels, namely "original research level," "evidence review level," "guideline and specification level," "practical knowledge level," "public education level," and "regulatory feedback level," and dynamically calls them according to applicability.
The "six - source evidence - based reasoning" paradigm (Source: Baichuan Intelligence)
The official evaluation results show that based on this "evidence - based reasoning" paradigm, the hallucination rate of Baichuan - M2 Plus's medical responses has significantly decreased. "Its credibility in multi - scenario evaluations has reached the level of senior clinical experts, and its ability to apply medical knowledge also surpasses that of top human doctors."
Imagine, if such a logically rigorous "evidence - based reasoning" paradigm is applied to more "serious content production" industries, such as education, can it bring more vertical large - model products comparable to the level of "business experts" in these industries?
Players in the vertical large - model field are continuously striving for technological iteration, and players in the general large - model field are also getting more involved in the technology and applications of vertical large models. In the newly released DeepSeek V3.2 by DeepSeek, in addition to the regular V3.2 version for the general public, there is also a V3.2 - Speciale version (only accessible via API service).
It is reported that DeepSeek - V3.2 - Speciale integrates the theorem - proving ability of long - term thinking and DeepSeek - Math - V2. It focuses on exploring the reasoning boundaries of the model and has stronger instruction - following, mathematical - proving, and logical - verification abilities. It won gold medals in top - level international competitions such as IMO 2025, CMO 2025, ICPC World Finals 2025, and IOI 2025. Its ICPC result is equivalent to "the second place among human contestants."
(Image source: DeepSeek)
To some extent, isn't DeepSeek - V3.2 - Speciale a "vertical large - model" product technology launched by DeepSeek for the fields of reasoning calculation and theorem proving?
Vertical large models will also enter a new stage of technological and product application development with the acceleration of their own technological iteration and the deeper involvement of general large - model