GuShiio.com 
Image source: Generated by AI
The 2025 Spring Festival holiday has just passed, but the shock wave caused by DeepSeek is still warm.
Through methods such as FP8 training, multi-word prediction, improved MOE architecture, multi-head latent attention mechanism (MLA), and SFT-free reinforcement learning, DeepSeek-V3 has surpassed the performance of top open source models and some closed source models such as Qwen2.5- 72B and Llama-3.1- 405B with extremely low training costs. DeepSeek-R1 has also shown an inference effect that surpasses OpenAI o1.
The success of the DeepSeek series of models has opened up a new path for the large model industry that was originally driven by computing power as the core logic, and has brought the world’s basic large models to a new level.
However, in addition to basic large models with “technical narrative” as the main theme such as DeepSeek, there is also a type of large model that deserves attention to, that is, AI technology innovation around core products and core scenarios.Application large model。
China has always been a major application country.
In 2024, as the supply of computing power is gradually catching up and the price of reasoning has dropped significantly, domestic AI applications have sprung up suddenly-whether it is Wenshengtu, Jimeng AI, Miaoya Camera, and Quick Hand Kerin in the field of Wenshengtu’s video, or Nano Search (formerly 360AI Search) and Tiangong AI Search in the field of AI search, Hoshino and Maobox in the field of AI companionship, or Bean Bao, Quark, Kimi, Tongyi, etc., all in 2024 ushered in an explosion in user volume.
These AI applications are inseparable from the support of the model capabilities behind them.For AI applications, application-oriented large models compete not on model parameters, but on application effects.
For example, the reason why Kimi was able to gain high attention in a short period of time was inseparable from the long text reading and parsing capabilities of the large model behind it; Quark’s 200 million users and 70 million monthly lives benefited from the “user-friendliness” of the large model behind it; Keling AI’s powerful literary video and graphic video functions rely on the support of Keling’s large model.
The evolution of basic large models is far from over, but as more and more companies begin to deploy AI applications in 2025, the development of application-oriented large models will be a necessary prerequisite for the comprehensive explosion of AI applications.
1. Why do big factories have more advantages in doing AI applications?
With the maturity and breakthrough of large model technology, the gradual improvement of computing power infrastructure, the continuous increase of national policies, the continuous emergence of killer applications such as Sora/Suno, and the strong growth of investment and financing in AI Agent/embodied intelligence/AI toys/AI glasses,2025 is the explosion year for AI applications, has almost become a broad consensus in the scientific and technological community.
And this consensus has also been accelerated by the popularity of DeepSeek. Because DeepSeek has pushed up the industry’s basic model capabilities, it has created a better development environment for AI applications.
According to “Jiazi Guangnian” observation, since the second half of 2024, well-known investment institutions such as Hillhouse Capital, Jingwei Ventures, Baidu Ventures, and Innocent have increased their investment in AI applications, especially for early projects targeting the AI application field. Betting; Some investors said that as of the end of 2024, the number of AI application projects actually funded in the primary market is at least twice as large as the number of projects actually announced.
Sensor Tower data also shows that in 2024, global mobile phone users spent US$1.27 billion on AI applications, and AI-related applications were downloaded 17 billion times on iOS and Google Play stores.
However, a cruel reality is that there are millions of AI applications, but only a few can truly maintain long-term operations, and very few can explode.
“Jiazi Guangnian” once reported on a website called “AI Cemetery”, which contained 738 AI applications that died or stopped running, including some former star projects: for example, the AI speech recognition product Whisper.ai launched by OpenAI, the well-known shell websites FreewayML and StockAI of Stable Diffusion, and the AI search engine Neeva, which was once regarded as a “Google competitor”(see “AI Cemetery, and 738 Dead AI Projects| Jiazi Light Year “).
So, what kind of AI application can run for a long time and be viable?
“Jiazi Guangnian” believes thatThe first is to focus on the model and give full play to the capabilities of the model;The second is to have sufficient insight into user needs.
Microsoft CEO Satya Nadella once said when looking forward to AI industry trends in 2025,”Applications centered on AI models will redefine various application areas in 2025”。In other words, the applications with fewer shell levels, closer to the model, and the more they maximize the model’s capabilities, the more attractive they will be for users to use and stay.
It is not difficult to observe the new list of AI products in January 2025 that among the top ten in the domestic list, 8 are AI assistant applications directly based on models.
Source: New list
To have strong enough insight into user needs, you must rely onhuge user base–Only by having enough users can user data and tags accumulate enough and thick enough, so that enterprises can dig out the most real pain points of users ‘needs.
These two points also mean that when doing AI applications, big manufacturers have more advantages.
Dachang has sufficient computing power and talents to develop its own models, so it can deploy AI applications directly on top of its own models without having to carry out layers of shells; Dachang also has a huge user base and mature traffic portal, which not only does user data It is richer and easier to mine needs, and also provides natural advantages for the promotion of AI applications; in addition, Dachang’s strong ecological integration capabilities also help provide products with richer functions and enhance the user stickiness of AI applications.
The aforementioned product list also proves this. Among the top ten applications, six are from major manufacturers.
In the latest interview with Zhu Xiaohu by Tencent Technology, Zhu Xiaohu also said that the data barriers of startups are not so high and are not suitable for making the underlying model, but need to grasp the “customers” more tightly on top of the underlying model. This also confirms the advantages of large factories in AI applications.
Overall, Dachang’sModels and applications are also causative of each other, and together constitute a growth flywheel:
The accumulation of data provided by the huge user base provides high-quality expectations for model development, helping to enhance model capabilities and make them better adapt to segmented scenarios and user needs; while the growth of model capabilities feeds back on applications, allowing applications to have stronger product power and attract more users.
We may be able to give it a name for this model with a large user base, driven by user needs, and better performance of capabilities in segmented scenarios.”Application big model”。The more AI applications are based on the “application model”, the more chance they have to successfully run out in theory.
For example, Quark, who ranks second only to DeepSeek on the list, is a typical representative.
“Jiazi Light Year” observed that in the recent battle between the gods in AI applications, quarks, which had rarely been mentioned before, were silently leading the way. The latest data from Analysys Analytics shows that at the end of 2024, Quark ranked first in mobile AI applications with 71.02 million monthly active users, surpassing the well-known Doubao and Kimi.
Source: Analysys Analysis
What deserves more attention is the “user stickiness” indicator.
According to statistics from third-party reports, Quark’s three-day Retention rate exceeds 40%. In comparison, the retention of high-profile bean buns and Kimi smart assistants on the market during the same period was about 25%; The “2024 Strong AI Product List” released by Qimai Data shows that Quark ranks first in the “Annual Strong AI Product App List” and the “Annual Product Download List”. Its cumulative downloads in 2024 exceed 370 million, ranking among all types of AI products, achieving a fault lead.
Among the many AI products on the list, Quark was not the first to launch a large model, but it quietly achieved a far lead in visits, downloads and user stickiness.quarkWhy can we break out in a highly competitive market?
Everything benefits from Quark’s “application first” product and model strategy.
2. Application first, backward upgrade of large models based on scenarios
Quark has focused on “intelligent and precise search” since the first day of searching. It not only relies on a concise ad-free interface and more accurate search results to quickly open a hole in the market,Based on the search business, vertical products such as Quark Network Disk, Quark Scan King, Quark Document, and Quark Learning have been developed around the student party and office worker groups. The scenes are gradually subdivided into learning and work fields.
Taking the field of learning as an example, in mid-2020, Quark launched the “Photo Search Question” function. During the epidemic, in response to the difficulties of many students being blocked from taking online classes at home and facing the inability to learn effectively, the Quark learning team upgraded the “photo search” function several times.
In the office field, Quark also launched a series of related functions such as text extraction, tables, handwriting removal, certificate scanning, and document format conversion based on the vertical scene of “scanning”.
The simple tool background, the increasingly rich scenario applications, and the initial innovation ecosystem with no advertising and no fees have allowed Quark’s user base to increase sharply, from one million to ten million, and the cumulative number of service users has exceeded 100 million.
In November 2023, Quark released the “Quark Big Model”, a large-scale model with hundreds of billions of parameters.
The Quark Big Model is a multi-modal large model independently developed by Quark based on the Transformer architecture. It trains and fine-tunes hundreds of millions of graphic data every day. It has the characteristics of low cost, high response, and strong comprehensive capabilities. Facing user needs and vertical classification scenarios of quark products, the quark large model pays more attention to practical applications, deriving vertical models such as general knowledge, medical care, and education to provide more professional and precise technical capabilities.
At the same time as the launch of the Quark Big Model, Quark upgraded the AI recognition effect of scanning products and the AI search capabilities of online disk products.
The first implementation scene of the Quark Big Model is health and medical care.
In December 2023, Quark announced a comprehensive upgrade of its health search function and launched the “Quark Health Assistant” AI application in December 2023. “Quark Health Assistant” combines medical knowledge mapping and generative dialogue capabilities to provide users with more comprehensive and accurate health information. It also supports users to conduct multiple rounds of questions and dialogues on health issues.
In January 2024, Quark successively launched functions such as “AI Learning Assistant”,”AI Listening”, and “AI PPT”. In July 2024, it launched a one-stop AI search-centered AI service on the mobile side. In August 2024, a new Quark PC with “system-level full-scene AI” capabilities was released.
For example, users search for “which attractions in Shanxi are derived from the black myth of Wukong.” The Quark Super Search Box integrates AI answers, original sources and historical searches-not only can it generate intelligent summaries like other AI searches, but also provides source display in the sidebar and uses AI search answers. The traditional search engine item-style web page presentation is retained. This improves the efficiency of users ‘information acquisition and also enhances the credibility of AI answers.
In addition, Quark has also built a one-stop information service system around the “super search box”, including intelligent tools such as network disks, scanning, document processing, and health assistants, realizing the full-process service from retrieval to creation and summary, to editing and storage, and sharing brings a seamless information service experience to users.
Different from many major manufacturers that imitate ChatGPT and launch the “All in One” Chatbo-type AI assistant,Quark’s strategy is “AI in All”-integrate AI capabilities into every aspect of the product and implement them in specific application scenarios.
From the initial photo search, to college entrance examination consultation, to smart office assistance, Quark’s product evolution has always centered on the needs of users in specific scenarios. Since then, Quark has successively launched and updated AI search questions, AI academic search, AI tips and other functions to create differentiated AI applications around learning and office scenarios.
The development history of quark AI in the past year, drawing: Jiazi Light Year
Among them, the “AI Search” function upgraded in November 2024 is a typical representative that best embodies Quark’s AI capabilities.
In fact, as early as December 2023, Quark launched an AI lecture assistant. At that time, AI question teaching assistants still relied more on the “knowledge base” of the question bank. AI could only teach users to do the questions in the question bank. The upgraded AI search product has stronger “intelligence”. It can not only answer the original questions in the question library, but also face new questions and problems. The use of the large model “Thought Chain (CoT)” allows Quark AI search questions to present the problem-solving ideas and problem-solving steps in turn, providing users with more detailed content analysis and learning guidance.
Compared with similar search products, most of which rely on question banks and can only answer questions in the K12 field, Quark’s AI search products can not only answer new questions in the K12 field, but also answer professional questions for postgraduate entrance examinations, public examinations, and various qualification examinations. Users only need to take photos or take screenshots, and Quark can search for the corresponding questions and provide professional content in pictures, videos and AI answers step by step. In addition, Quark’s “AI Search Questions” can also provide answers to questions in sub-fields such as law and medicine.
Quark’s answer to the real question of the judicial examination
At the same time, Quark’s “AI Search Questions” can also use AI capabilities to provide in-depth explanations of the knowledge points and test points in the question, accurately locate key steps, allowing users to not only learn this question, but also learn this “from one example”. A class of questions.
Quark’s powerful ability to “AI search questions” not only relies on Quark’s years of searching, the accumulation of enough high-quality materials and user needs in learning scenarios, but is also inseparable from Quark’s launch during the same period.Gnostic learning modelsupport.
The “Gnostic” Big Model is trained by the Quark Technology Team on the basis of the “Quark Big Model” and through high-quality data accumulated through years of deep cultivation in the field of education. It not only has the thinking chain capabilities that many top models have, but also It can transform the thinking process into a language that students can understand and is more in line with their learning process.
In other words, they are also explaining a problem to students. The “Gnostic” model knows more about which knowledge points to explain and how to build problem-solving ideas.
Take the 2024 Beijing College Entrance Examination mathematics question as an example, enter it into DeepSeek and Quark respectively, and the answers are as follows:
Answer given by DeepSeek
Quark’s answer
It can be seen that compared with DeepSeek’s lengthy chain of thoughts with the official and detailed answers, Quark’s answer is more concise, more like explaining a question.
Due to the large number of “knowledge explanation” and “popular science” scenarios, the education industry has put forward high requirements for the multimodal capabilities of models. However, existing multimodal models have poor recognition of formulas, handwritten notes, etc., especially the fine-grained understanding of graphics.
In order to solve this problem, the quark “Gnostic” large model builds a large-scale domain professional training corpus through a large-scale multimodal pre-training base, and at the same time ensures better understanding in terms of model structure.
In the latest evaluation, the accuracy and score rate of Quark’s “Gnostic” learning model on postgraduate mathematics questions 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, Quark’s correct rate and scoring rate are also in an absolute leading position.
Display of mathematical evaluation results of the “Gnostic” model Source: Quark
Unlike companies such as DeepSeek, which have the ability to develop purely basic models, the Quark development model is user-oriented.Taking AI writing as an example, the Quark technology team used multi-stage CoT and search enhancement technology to develop a Quark cultural creation model that can generate long texts of more than 8000 words in response to the needs of young Quark users to write reports and papers., ensuring the number of words follows. Even DeepSeek can currently only generate articles of up to 3000 words.
In addition, Quark’s AI writing function is equivalent to an “online text editor”. Users can perform complex operations such as deleting, polishing, and expanding the generated articles, which is also inseparable from the ability of Quark’s cultural creation model. support.
It can be said thatWhile the world is “rolling” large model parameters, Quark has focused more on practical application scenarios, and directed upgrades and optimizes model capabilities based on user needs.As of now, Quark has formed system-level full-scene AI capabilities.
Source: Quark
3. Ali AI To C accelerates
As one of Ali’s four strategic innovation businesses, Quark’s every move represents not only himself, but also the direction of the entire Ali AI To C business.
On January 15, Quark upgraded its brand Slogan-“AI All-round Assistant for 200 Million People”, highlighting 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 Quark and other AI To C businesses are located.
Recently, Ali has taken frequent actions in the field of AI To C: first, Wu Jia, a “young and strong” executive, returned to Alibaba Group to explore the AI To C business; then Ali’s AI application “Tongyi” was officially spun off from Alibaba Cloud and merged into the Ali Intelligent Information Business Group; and recently, according to media reports, Tmall Elf’s hardware team is also working integrating with Quark product team. Its focus includes the planning and definition of new generation AI products and the integration with Quark AI capabilities. After the team integrates, the new team will also explore new hardware directions including AI glasses.
Since then,Quark, Tongyi App, and Tmall Elf will serve as productivity tools, Chatbot, and AI hardware respectively to provide differentiated services to users.
On February 6, Ali’s ToC field welcomed a heavyweight figure-Professor Steven Hoi, the world’s top artificial intelligence scientist, officially joined Alibaba as vice president of Alibaba Group, reporting to Wu Jia, responsible for the multi-modal 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 multimodal basic models of AI To C services and basic research and application solutions related to Agents, greatly enhancing the end-to-end closed-loop capabilities of Alibaba’s AI application C-end products in model-combined applications. Transition. Once the capabilities of multimodal basic models have achieved breakthroughs, quark and other C-terminal applications will have new space for exploration in business.
At the same time, Ali’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 have analyzed that at the beginning of 2025, the joining of world-class top scientists can be regarded as an important signal for Ali AI To C to increase its investment in talents and resources. The top talent team of the large model will support Ali AI To C’s in-depth exploration in multi-modal Agents and other directions, and will also open up the imagination for the next stage of building a user-oriented AI application platform.
Today, Byte has invested heavily in the field of AI applications and restarted the “App Factory” strategy through vigorous streaming, internal horse racing, and active going to sea; Tencent has launched “Yuanbao” and “Yuanqi” in the direction of AI assistants and agents. Two products have regained public attention through the newly launched personal knowledge management tool ima.copilot; Baidu has launched an AI product matrix including Wenxinyiyan, Wenxinyige, Orange AI, Super Canvas, etc., using a “big and comprehensive” style to carry out “saturated attacks” on friends. In addition, new startups such as the big model “Six Tigers” and DeepSeek have also launched AI applications. Ali’s AI To C business can be said to be surrounded by strong enemies, and the pressure can be imagined.
However, if there are difficulties, there must be solutions.And with more than 200 million users and top monthly activity ranking. It also proves the correctness of Quark’s style of play and the bright future of Ali’s AI To C business.
At a time when AI technology has entered the deep application area, Quark’s innovative paradigm has given us key enlightenment:Real technological advancement lies not only in how many technological peaks you climb, but also in how many scientific and technological achievements you can transform into value that can be touched at the fingertips of users.Only when users truly make choices and use practical actions to vote for AI applications, this breakthrough battle related to the practical application of AI technology may come to the real competition point that determines the future industrial landscape.
