Picture source: Generated by Unbounded AI

The Spring Festival holiday in 2025 has just passed, but the shock wave caused by DeepSeek still remains undissipated.

Through FP8 training, multi-word meta prediction, improved MOE architecture, multi-head potential attention mechanism (MLA), SFT-free reinforcement learning and other methods, DeepSeek-V3 has surpassed Qwen2 with extremely low training costs. With the performance of top open source models such as 5-72B and Llama-3.1-405B, DeepSeek-R1 has shown inference effects that surpass OpenAI o1.

The success of the DeepSeek series models has opened up a new path for the big model industry that was originally driven by computing power as the core logic, and has taken the global basic big model to a new level.

However, in addition to basic large models such as DeepSeek with the theme of "technical narrative", there is another type of large models that are worth paying attention to, that is, AI technology innovation around core products and core scenarios. big application model.

China has always been a major application country.

In 2024, against the backdrop of the gradual catching up with computing power supply and the sharp decline in inference prices, domestic AI applications emerged - whether it is the Zhimeng AI in the fields of Wensheng Pictures and Wensheng Video, and Miaoya Camera, Kuaishou is also a nano-search in the field of AI search (formerly 360AI search), Tiangong AI search, Hoshino and cat boxes in the field of AI companionship, or a bean bag, quark, Kimi, Tongyi, etc., all of which are all 2024 ushered in an explosion in user volume.

The AI ​​applications are inseparable from the support of the model capabilities behind them. For AI applications, the application-based big models are not the model parameters, but the application effects.

For example, the reason why Kimi was able to gain high attention in a short period of time was closely related to the long text reading and parsing capabilities of the big model behind it; the 200 million users of Quark and 70 million monthly active users Thanks to the "user-friendliness" of the quark model behind it; the powerful Wensheng Video and Tusheng Video functions of Keling AI rely on the support of the Keling model.

The evolution of basic big models is far from over, but as more and more companies begin to deploy AI applications in 2025, the development of application-based big models will be a comprehensive explosion in conjunction with AI applications. A necessary prerequisite.

1. Why are big companies more advantageous in AI applications?

With the maturity and breakthrough of large model technology, the increasing computing power infrastructure Improvement, the continuous increase in national policies, the continuous emergence of killer applications such as Sora/Suno, and the strong growth in investment and financing in the fields of AI Agent/embodied intelligence/AI toys/AI glasses. 2025 is the year of explosion in AI applications, and it has almost been becomeBroad consensus in the science and technology community.

And this consensus has also accelerated due to the popularity of DeepSeek. Because DeepSeek pushes up the water level of the industry's basic model capabilities, it creates a better development environment for AI applications.

According to the observation of "Jiazi Light Year", from the second half of 2024 to the present, well-known investment institutions such as Hillhouse Capital, Jingwei Venture Capital, Baidu Venture Capital, Innolux and other well-known investment institutions have increased their investment in AI applications, especially Betting on early-stage project bets in the field of AI applications; some investors said that as of the end of 2024, the number of AI application projects that have been truly financed in the primary market was at least twice as many as the number of projects actually announced.

Sensor Tower data also shows that in 2024, global mobile phone users spent $1.27 billion on AI applications, and AI-related applications were downloaded 17 billion times in iOS and Google Play stores.

However, a cruel reality is that there are thousands of AI applications, only a few can truly maintain long-term operation, and very few can become popular.

"Jiazi Light Year" once reported on a website called "AI Cemetery", which contains 738 AI applications that have died or stopped running, including some of the former celebrity projects: such as OpenAI launches Whisper.ai, the well-known AI voice recognition product of Stable Diffusion, FreewayML, StockAI, and Neeva, an AI search engine that was once regarded as a "Google competitor" (see "AI Cemetery," and 738 Dead AIs for details Project | Jiazi Light Years》).

So, what kind of AI application can last for a long time and be vital?

"Jiazi Light Year" believes that, first, we must take the model as the core and give full play to the model's capabilities; second, we must have strong enough user needs insight.

Microsoft CEO Satya Nadella once said when looking forward to the trends in the AI ​​industry in 2025 that "applications with AI models as the core will redefine various application fields in 2025." In other words, the fewer the shell levels, the closer the model is to the model, and the more it can maximize the model's capabilities, the more it can attract users' use and stay.

Observing the new list of AI products in January 2025 is not difficult to find that among the top ten domestic lists, 8 are directly built on models and AI assistants. application.

Photo source: Xinbang

To have strong enough insight into user needs, you must rely on a huge user base - only with enough users, user data and tags Only by accumulating enough and thick enough can enterprises tap into the most real pain points of users' needs.

These two points also mean that large companies have more advantages in AI applications.

The large manufacturers have sufficient computing power and talents to develop their own models, so they can be covered without the need for layers of shells.Directly deploy AI applications on the self-developed model; large manufacturers also have a huge user base and mature traffic entrances, which not only enrich user data and easier to explore demand, but also provide natural advantages for the promotion of AI applications; in addition, , the strong ecological integration capabilities of large manufacturers also help to provide more diverse functions for products and enhance user stickiness of AI applications.

The aforementioned product list also proves that it is hungry. Six of the top ten applications come from major manufacturers.

In the previous interview with Zhu Xiaohu by Tencent Technology, Zhu Xiaohu also said that the data barriers of startups are not that high and are not suitable for making underlying models. Instead, they need to capture "customers" on the underlying models. Tighter. This also indirectly confirms the advantages of large manufacturers in AI applications.

From the overall perspective, the models and applications of large manufacturers are also causal and effective, and together form a growth flywheel:

The data accumulation provided by the huge user base provides model research and development. High-quality predictions help enhance model capabilities and better adapt to segmented scenarios and user needs; while the growth of model capabilities feeds back to applications, allowing applications to have stronger product power and attract more users.

This model with a large-scale user base, driven by user needs, and better ability in segmented scenarios, can be named "application model". The more AI applications are based on the "application big model", the more chance they will have will be successful.

For example, the quark, which ranks second only to DeepSeek in the list, is a typical representative of it.

"Jiazi Light Year" observed that in the recent melee of the gods of AI applications, the quarks, which were rarely mentioned before, are silently leading the way. The latest data from iResearch Analysis shows that by the end of 2024, Quark ranked first in the mobile AI application list with 71.02 million monthly active users, surpassing the well-known Doubao and Kimi.

Photo source: iResearch Analysis

What is more worthy of attention is the "user stickiness" indicator.

According to third-party reports, the three-day retention rate of quarks exceeds 40%. In contrast, the retention rate of highly anticipated bean bags and Kimi smart assistants in the market during the same period was about 25%; Qimai data The "2024 Annual Strength AI Product List" released shows that Quark ranks first in the "Annual Strength AI Product App List" and "Annual Product Download List". Its cumulative download volume in 2024 exceeded 370 million, among which various AI types It stands out among the products and achieves fault-type leadership.

Among the many AI products on the list, Quark was not the first to launch a big model, but it quietly achieved a far-reaching lead in visits, downloads and user stickiness. Why can Quark emerge in a highly competitive market?

It all comes from Quark's "application-first" product and model strategy.

2. Application first, reverse the scene-based upgrade of the big model

Quark has focused on "intelligent and accurate search" since the first day of search. It not only quickly breaks a hole in the market with a simple and ad-free interface and more accurate search results, but also based on search business. Nearby products such as Quark Netdisk, Quark Scan King, Quark Documents, Quark Learning have been derived around the student party and office workers, and the scenes are gradually subdivided into the fields of learning and work.

Taking the field of learning as an example, in mid-2020, Quark launched the "photo search" function. During the epidemic, in response to the difficulties of many students being blocked from their homes and facing the difficulties of not being able to learn effectively, the Quark Learning Team has upgraded the "photo-taking and question-search" function many times. ‍

In the office field, Quark has also launched a series of related functions such as extracting text, tables, removing handwriting, document scanning, and document format conversion from the vertical scene.

Simple tool background, increasingly rich scenario applications, and the initial new ecosystem without advertising and no fees have allowed Quark's users to soar, from one million to tens of millions, with cumulative services The number of users exceeds 100 million.

In November 2023, Quark released the "Quark Big Model" a 100 billion parameter model.

The Quark big model is a multimodal big model developed by Quark based on Transformer architecture. It trains and fine-tunes of graphic data every day, and has low cost, high response and strong comprehensive capabilities. and other characteristics. In response to user needs and vertical scenarios of quark products, the quark model focuses more on practical applications and derives vertical models such as general knowledge, medical care, and education to provide more professional and accurate technical capabilities.

At the same time as the launch of the quark model, Quark upgraded the AI ​​recognition effect of scanning products and the AI ​​search capabilities of network disk products.

The first landing scenario for the quark model is health and medical care.

In December 2023, Quark announced the comprehensive upgrade of its health search function and launched the "Quark Health Assistant" AI application in December 2023. "Quark Health Assistant" integrates medical knowledge graphs and generative dialogue capabilities, providing users with more comprehensive and accurate health information, and also supports users to conduct multiple rounds of questions and conversations on health issues.

In January 2024, Quark successively launched functions such as "AI Learning Assistant", "AI Listening and Memory", and launched AI Search on mobile in July 2024. The center’s one-stop AI service was released in August 2024. The new quark PC end with the ability of “system-level full-scene AI”.

For example, users search for "Which attractions in Shanxi based on the Black Myth Wukong". The Quark Super Search Box integrates AI answers, original source and historical searches—not only can generate intelligent summary like other AI searches, but also provides source display in the sidebar and search in AI The traditional search engine entry-style web page presentation is retained under the answer. This improves the user's information acquisition efficiency and enhances the AI ​​answersTrustworthiness.

In addition, Quark has built a one-stop information service system around the "super search box", including intelligent tools such as network disk, scanning, document processing, and health assistant, realizing the transformation from search to creation and summary. , and then to the full-process services of editing, storage and sharing, bringing users a seamless information service experience.

Unlike many Chatbot-like AI assistants that imitated ChatGPT and launched "All in One", Quark's strategy is "AI in All" - integrating AI capabilities into every link of the product and implementing it To specific application scenarios.

From the initial photo search, to college entrance examination application consultation, and then to intelligent office assistance, Quark's product evolution has always revolved around user needs in specific scenarios. Since then, Quark has successively launched and updated functions such as AI question search, AI academic search, and AI tips to create differentiated AI applications around learning and office scenarios.

The development history of Quark AI in the past year, mapped by: Jiazi Light Year

Among them, the "AI Question Search" function upgraded in November 2024 is the most concentrated reflection of Quarks. A typical representative of AI capabilities.

In fact, as early as December 2023, Quark launched the AI ​​question-telling assistant. At that time, AI question-teaching assistants relied more on the question bank as the "knowledge base", and AI could only teach users how to do the questions in the question bank. The upgraded AI question search product has stronger "intelligence", which can not only solve the original questions in the question bank, but also face new and difficult questions. The use of the big model "Thinking Chain (CoT)" allows Quark AI to search questions to present the problem-solving ideas and steps to solve the problem in sequence, providing users with more detailed content analysis and learning guidance.

Compared with similar question search products, most of which rely on question banks and can only answer questions in the K12 field, Quark's AI question search products can not only answer new questions in the K12 field, but also answer postgraduate entrance examinations, public examinations, and various Professional questions for the qualification certificate examination. Users only need to take photos or take screenshots, and Quark can search for the corresponding questions and give professional content in pictures, texts, videos and AI answers in steps. In addition, Quark's "AI Search Questions" can also give answers to questions in sub-sectors such as law and medicine.

Quak's answer to the real questions of the judicial examination

At the same time, Quak's "AI Question Search" can also use AI capabilities to provide in-depth explanations of the knowledge points and test points in the questions and accurately position them. The key step allows users to not only learn this question, but also learn this kind of question by "learning one by one and applying it to others".

Quark's powerful ability to "AI search questions" is not only based on Quark's years of search precipitation and sufficient high-quality materials and user needs accumulated in the learning scenario, but also inseparable from Quark in the same period. The "Genzhi" learning model launched is supported.

The "Genzhi" big model is trained by the Quark technical team based on the "Quark Big Model" through many years of hard-working and accumulating high-quality data in the field of education. It not only has many top modelsThe thinking chain ability that everyone has is able to transform the thinking process into a language that students can understand and more in line with their learning process.

In other words, it is also about explaining a question to students. The "Lingzhi" big model knows more about what knowledge points to explain and how to build problem-solving ideas.

Take the 2024 Beijing College Entrance Examination Mathematics Questions as an example, enter them into DeepSeek and Quark respectively. The answers obtained are as follows:

< /p>

DeepSeek's answer

Quark's answer

You can see that comparing DeepSeek's long-standing thinking chain narrative and official and detailed answers, Quark's answer is more concise, more like explaining a question.

The education industry has put forward high requirements for the multimodal ability of the model due to a large number of "knowledge explanation" and "popular science" scenarios. However, existing multimodal models have poor recognition of formulas, handwritten notes, etc., especially the fine-grained understanding of the graphics is relatively poor.

In order to solve this problem, the Quark "Glory" big model built a large-scale field professional training corpus through large-scale multimodal pre-training base, and at the same time, ensuring more in the model structure. Good understanding effect.

In the latest review, the accuracy and scoring rate of Quark's "Lingzhi" learning model in postgraduate entrance examination math problems can already be comparable to OpenAI-o1, and far exceed other domestic models. In many important tests such as domestic mathematics competitions and college entrance examinations, the accuracy and scoring rate of quarks are also in an absolute leading position.

The mathematical evaluation results of the "Lingzhi" big model are displayed. Source: Quark

Unlike the pure basic model capabilities of companies such as DeepSeek, the quark R&D model is oriented towards user needs. . Taking AI writing as an example, the Quark technical team used multi-stage CoT and search enhancement technology to develop a quark cultural and creative model that can generate long articles of more than 8,000 words in response to the needs of young quark users to write reports and papers. Ensure the word count is in line with the effect. Even DeepSeek can only generate articles of up to 3,000 words at present.

In addition, Quark's AI writing function is equivalent to a "text online editor", where users can delete, polish, and expand the generated articles, and this is also inseparable from it. Support of Quark's cultural and creative model capabilities.

It can be said that when the world is "volume" big model parameters, Quark has focused more on practical application scenarios and targeted upgrades and optimizes model capabilities based on user needs. As of now, Quark has formed system-level full-scene AI capabilities.

Photo source: Quark

3.Ali AI To C Acceleration

AsOne of Alibaba's four major strategic innovative businesses, Quark's every move represents not only itself, but also the direction of the entire Alibaba AI To C business.

On January 15, Quark upgraded its brand Slogan—“A AI All-round Assistant for 200 million people”, revealing a new business trend to accelerate the exploration of AI To C applications. Recently, Alibaba founder Jack Ma suddenly "flashed" in Alibaba's Hangzhou Park and also went to the office area where AI To C businesses such as Quark are located.

In recent times, Alibaba has been making frequent moves in the field of AI To C: First, Wu Jia, the senior executive of the "Young and Strong School", returned to Alibaba Group to explore the AI ​​To C business; then Alibaba's AI application "Tongyi" is officially Split from Alibaba Cloud and merged into Alibaba Intelligent Information Business Group; and according to media reports, Tmall Ghost’s hardware team is currently working with the Quark product team, and its focus includes the planning and definition of the new generation of AI products, as well as Fusion with quark AI capabilities. After the team is integrated, the new team will also explore new hardware directions including AI glasses.

From now on, Quark, Tongyi App, and Tmall Genie will serve as productivity tools, Chatbot, and AI hardware respectively to provide users with differentiated services.

On February 6, Alibaba's ToC field welcomed a heavyweight figure - Professor Xu Zhuhong, a world-leading artificial intelligence scientist, officially joined Alibaba and became the vice president of Alibaba Group. , report to Wu Jia, responsible for the multimodal basic model of AI To C business and basic research and application solutions related to Agents.

According to insiders, Professor Xu Zhuhong will focus on the multimodal basic model of AI To C business and Agents-related basic research and application solutions, greatly improving the model combination application of Alibaba AI application C-end products end-to-end closed-loop capability transition. Once the capabilities of multimodal basic models have been made, C-end applications such as quarks have new room for exploration in their business.

At the same time, Alibaba's AI To C business is forming a top AI algorithm research and engineering team to attract a large number of outstanding talents in the industry to join. Some industry insiders analyzed that the franchise of world-class top scientists at the beginning of 2025 can be regarded as an important signal for Alibaba AI To C to increase its investment in talents and resources. The top talent team of big models will support Alibaba AI To C's in-depth exploration in multimodal Agents and other directions, and will also open up imagination space for building a user-oriented AI application platform in the next stage.

Now, Byte has put a heavy bet in the field of AI applications, restarted the "App factory" strategy through vigorous investment, internal horse racing, and active overseas trips; Tencent has launched the "Yuanbao" in the direction of AI assistants and agents. ", "Compens" two products, and have regained public attention through the latest personal knowledge management tool ima.copilot; Baidu has launched Wen Xin Yiyan, Wen Xin Yige, and Orange AThe AI ​​product matrix including I, super canvas, etc. uses a "big and complete" approach to conduct "saturated attacks" on friendly companies. In addition, new startups such as the big model "Six Little Tigers" and DeepSeek are also focusing on AI applications. Alibaba's AI To C business can be described as strong enemies, and the pressure can be imagined.

However, if there are difficulties, there must be solutions. Through the "AI in All" strategy and precise control of user needs, Quark proves that it can achieve strong product strength without competing with parameters, relying on "application big models" and accurate grasp of user needs. This is another The version is "low cost and high efficiency"; and the number of users exceeds 200 million and the monthly active users ranking top. It also proves the correctness of Quark's strategy and the bright future of Alibaba's AI To C business.

At the moment when AI technology enters the "deep water zone of application", Quark's innovation paradigm has given us key inspiration: the true technological advancement lies not only in how many technological peaks it can climb, but also in how many scientific and technological achievements it can transform. Value that can be touched by users' fingertips. Only when users really make choices and vote for AI applications with practical actions, this breakthrough battle related to the practical use of AI technology may come to the real match point that determines the future industrial structure.