Written by Yi Si

Edited by Wang Yisu

After a small-scale test in 2023, the three major operators are full steam ahead with large models in 2024.

In the just-passed May, China Telecom, China Mobile, and China Unicom, the three major operators, collectively completed a new round of large model evolution, especially around the research and development and application of large models, showcasing their latest achievements in technological innovation and industrial upgrading.

Rapid evolution stems from the strategic layout of the three major operators in AI.

China Telecom stated in its 2023 annual report that the company will seize the trend of a new round of technological revolution and industrial transformation represented by artificial intelligence in 2024; China Mobile has also publicly stated that the Jiutian AI team will complete multi-modal large model training by the end of 2024, comparable to GPT-4, including 8 sub-platforms, 287 core algorithmic capabilities, and 37 large-scale applications.

Zhang Kai, a consultant from Nomura Research Institute, believes that operators' basic communication business and public market cannot generate large-scale growth in the short term, and ICT business and government and enterprise market are new business growth spaces. With the acceleration of digitalization, AI will inevitably become one of the social infrastructures. Operators' attempt to develop large model basic capabilities at this time is a strategic choice.

Intelligence has been a technological innovation direction that operators have emphasized in recent years, and AI functions have been added to various products under their brands, especially network products on which they rely for survival. As early as 2020, a survey data released by Omdia's ICT-Enterprise Insights showed that nearly 80% of new operators believed that using AI technology was an "important" or "very important" IT project, with nearly 60% of operators planning to increase investment in AI tools.

It is worth noting that operators have formulated different strategies from leading cloud vendors when it comes to deploying large models: they are not focusing on general large models, but rather on industry-specific large models to accelerate their implementation.

In terms of technology and products, operators find it difficult to compete with cloud vendors and large model companies with a technology company background, but for cloud vendors who have been deeply rooted in the digital market for many years, 5G's sharp edge and deep cultivation in the government and enterprise market have made them heavyweight players in the AI market that are difficult to ignore.

For cloud vendors who have high hopes for large models, these competitors are both to be respected and to cooperate with.

Operators' Large Models, Fast Acceleration and Strong Momentum

After a small-scale test in July 2023, operators have significantly accelerated the deployment speed of large models this year.

In the just-passed May, the three major operators collectively completed a new round of large model evolution, especially around the research and development and application of large models, showcasing their latest achievements in technological innovation and industrial upgrading.

China Mobile can be said to be the earliest operator to enter the large model market. In 2013, China Mobile established the "Jiutian" AI team. Since then, China Mobile has been surging ahead. In 2023, two industry-specific large models, "Jiutian·Haisuan Government Affairs" and "Jiutian·Customer Service," were released, taking the lead in implementation in industry scenarios.

In October of the same year, China Mobile jointly released the "Jiutian·Zhongqing" base large model with multiple central enterprises in the fields of telecommunications, energy, aviation, etc. Its training data scale exceeds two trillion Tokens, integrating expertise from eight major industries including telecommunications, energy, steel, construction, and transportation, and possessing five major advantages such as independent technological breakthroughs, safety, reliability, and controllability.

At the just-concluded Digital China Summit, China Mobile unleashed three major moves:

Firstly, the "Jiutian" intelligent base is composed of WanCard computing power, a trillion-parameter model, and the Baihui platform; secondly, it opens three major AI bases, including a large model training base, a large model evaluation base, and a large model industry innovation base, providing ultra-large-scale, ultra-long-term stability, and ultra-high-efficiency large model training services to the entire society. Thirdly, on the product application side, China Mobile launched 23 AI+ products and 20 AI+DICT industry applications. For example, 5G+AI new calls support real-time transcription and subtitle translation during calls; video ringtones introduce new functions such as text-to-image and text-to-music.

China Unicom, entering the large model race with the "Honghu Graphic and Text Large Model," positioned its first move in text-to-image, which is also the first AI large model for operators' value-added services. Compared to last year, China Unicom has been significantly faster in 2024. At the Mobile World Congress in February this year, China Unicom released the Yuanjing "1+1+M" large model system, consisting of one basic large model, one large model base, and M industry-specific large models, to solidify its core technological capabilities in the large model competition.

It is worth noting that at the model level, compared to the other two operators, China Unicom can provide models with different parameter volumes. Among them, the Honghu Graphic and Text Model comes in two versions with 800 million and 2 billion parameters. In terms of basic models, China Unicom has trained models with different parameter volumes such as 1 billion, 7 billion, 13 billion, 34 billion, and 70 billion by drawing on the laws of animal intelligence evolution.

From China Unicom's newly released Yuanjing Large Model MaaS platform and Yuanjing Economic Large Model, it is clear that it has accelerated the speed of popularizing large models across the entire industry.

Among them, the Yuanjing Large Model MaaS platform can provide three major services: a model library, a toolbox, and a native app store. The model library has built-in over 150 models, including general large models, industrial domain large models, and basic small models; the toolbox has built-in over 100 ready-to-use tools, including model selection, modification, and usage, enabling rapid customization with zero coding and low thresholds.

The Yuanjing Economic Large Model is mainly targeted at various key scenarios in the economic field, such as macroeconomics, industrial economics, industry analysis, enterprise production and operation digitization, business environment, and government services, developing special functions such as intelligent question answering, intelligent analysis, forecasting, and early warning, providing economic knowledge integration, analysis, and output services through human-machine dialogue interactions.

Compared to the first two, China Telecom, which has the highest market share in cloud services, has a more systematic strategy, with particular emphasis on the construction of model basic capabilities. China Telecom proposed the "1+1+1+M+N" strategy very early on, namely, one intelligent computing cloud base, one general large model base, one data base, M internal large models, and N industry-specific large models.

Based on this strategy, China Telecom initiated the development of a semantic large model in December 2022 and released the Chengxin Semantic Large Model with 100 billion parameters in 2023, successively open-sourcing 7B, 12B, and 52B large models.

In May of this year, China Telecom launched the Chengxin Speech Large Model and announced its full open-sourcing. It is understood that the model can simultaneously recognize and understand over 30 dialects, including Cantonese, Shanghainese, Sichuanese, and Wenzhou dialect, making it the largest speech recognition large model in China that supports the most dialects. At the same time, the release of this model has also enabled China Telecom to form a complete large model layout spanning semantics, speech, vision, and multi-modality.

Commercialization and Implementation, Operators Converging on the Same Path

Entering 2024, large models have come to a critical period of commercialization and implementation.

After a year of honing, it is no surprise that, like the development trend of the large model industry, operators have also shifted from large model training and research and development to how to implement large models. However, as operators, they have converged on the same path in terms of path selection.

Optical Cone Intelligence observed that "general large models + industry-specific large models + large model platforms + AI innovative products" is the basic operation of operators deploying large models.

Before the troops move, provisions must be made. To train high-quality large models and meet reasoning demands, efficient computing power is indispensable. Recently, the three major operators have been continuously adding computing power to the racetrack.

Recently, China Mobile issued a procurement tender announcement, initiating the procurement of new intelligent computing centers from 2024 to 2025, planning to procure up to 7,994 AI servers, becoming the largest domestic AI server centralized procurement. China Telecom procured 4,175 AI computing servers last August. China Unicom also procured 2,503 AI servers for the first time at the group level.

In addition to bidding, operators are also accelerating the construction of intelligent computing centers in various regions across the country. Among them, China Mobile has successively put into operation multiple intelligent computing centers distributed in Beijing-Tianjin-Hebei, the Yangtze River Delta, Guangdong-Hong Kong-Macao, Hubei, Chengdu-Chongqing, and Guizhou. At this year's Digital China Summit, China Mobile announced that it would build a new intelligent computing network cluster by the end of the year, putting into operation three near-20,000-card ultra-large-scale individual intelligent computing centers and 12 regional intelligent computing centers, extensively upgrading 1,500 edge nodes, and increasing the ratio of training/inference computing power resources to 1:10.

Tianyun, a subsidiary of China Telecom, is also moderately advancing the construction of intelligent computing infrastructure. Currently, it has planned and built a national "2+3+7+X" public intelligent computing cloud pool, building two 10,000-card intelligent computing clusters in Beijing-Tianjin-Hebei and the Yangtze River Delta regions and creating large-scale green intelligent computing pools in the western region. And as early as last year, China Telecom already had over 1,000-card training resources in nodes in Beijing, Shanghai, Jiangsu, Guizhou, Ningxia, and Inner Mongolia.

China Unicom has deployed a multi-level computing system with hotspot intensification, cross-regional radiation, and edge coverage, building a "1+N+X" national hierarchical intelligent computing resource with over 400,000 IDC racks and cloud pools covering more than 230 cities, achieving diversified computing power supply in an integrated manner.

To realize the implementation of large model technology, the three major operators have built model platforms (MaaS platforms).

China Mobile's convergence platform gathers and opens over 30 "proprietary + open-source + ecosystem" large models, over 20 high-quality datasets, and includes a full chain of large model development tools covering model experience, tuning, deployment, reasoning, and evaluation. China Telecom launched the Chengxin MaaS platform, integrating self-developed models and mainstream ecological partner basic large models, providing end-to-end large model solutions from "evaluation-selection-deployment-data-purchase-training-inference-usage." China Unicom released the Yuanjing Large Model MaaS platform, providing model library services, toolbox services, and native app store services.

Looking at a series of measures taken by operators, whether it is general models, intelligent computing centers, or MaaS platforms, their ultimate goal is to achieve rapid implementation by building multiple deep and vertical industry-specific large models.

Some have defined general models and industry models in this way: general large models are the "model room" in large models, while industry-specific large models are "commercial houses" that can directly face users.

"The three major telecommunications operators have avoided the general large model track, combined their data and user advantages, utilized the technological changes brought by large models, selectively implemented industry AIGC, made existing services 'intelligent,' and better served users and customers, which is a wise choice," said Fu Liang, a telecommunications analyst.

Based on the Jiutian large model, China Mobile has launched over 20 industry-specific large models targeting areas such as government governance, industrial production, people's livelihood services, and communications characteristics. In addition, based on global general knowledge, China Mobile has also conducted enhanced training on professional knowledge in industries such as telecommunications, power, transportation, and energy.

Based on the basic model and base, China Unicom has launched over a dozen industry-specific large models covering equipment manufacturing, warehousing logistics, power, steel, and other industries, and has already achieved implementation and application in industries such as networks, customer service, anti-fraud, industry, and government affairs, achieving phased results.

China Telecom has gone deeper in terms of industry-specific large models, first launching an education large model and subsequently releasing 12 industry-specific large models such as government services, grassroots governance, and health insurance. This year, China Telecom has further segmented specific industries, launching eight industry-specific sub-models including the campus help desk large model, mental health large model, emergency warning vision large model, Yizhiyu flood prevention warning auxiliary large model, and cultural heritage explanation large model.

Optical Cone Intelligence observed that operators all focus on the government and enterprise field in industry-specific large models, and their determination to deploy is far greater than people imagine. However, to achieve rapid implementation, it is not only necessary to study core industries but also to solve problems in segmented scenarios. In this regard, operators have shown unprecedented tacit understanding, and the first area to be implemented has pointed to - intelligent customer service.

Optical Cone Intelligence learned that in 2023, China Mobile formed a special team for the customer service industry large model. In April of the same year, China Mobile's customer service large model algorithm passed the service provider filing; in December, the customer service large model technology achieved industry application breakthroughs and conducted internal testing with over 20,000 employees.

China Unicom's Yuanjing Customer Service Large Model helped reduce average call duration by 28 seconds and increase the first-time problem resolution rate by 1.6pp. At the same time, based on the customer service large model, China Unicom has also created 26 AI innovative services. Through full-process, full-scenario digital empowerment, it comprehensively improves customer perception, front-line efficiency, and management effectiveness, achieving an intelligent service ratio of over 85%, reducing average call duration by 15%, and achieving a user satisfaction rate of over 98%.

China Telecom's newly released Chengxin Speech Recognition Large Model not only technically supports mixed speech in 30 dialects but has also been pilot-applied in the intelligent customer service systems of Wanhao in Beijing, Fujian, Jiangxi, Guangxi, and other regions, handling about 2 million calls daily. At the same time, the intelligent customer service Yisheng platform has also integrated the speech understanding and analysis capabilities of the Chengxin large model, achieving daily processing of 1.25 million customer service calls.

As everyone knows, customer service has always been the main business direction of operators and an industry with a high degree of intelligence penetration.

Based on years of deep cultivation in the customer service field, the three major operators have accumulated platform capabilities, rich experience, and vast amounts of data, providing fertile soil for the training of customer service large models. At the same time, with the support of large model technology, operators continuously enhance the capabilities of intelligent customer service and related products by leveraging the advantages of large models such as multi-modality and anthropomorphic interaction, greatly reducing human customer service response time and assisting human thinking, improving user experience.

Entering the era of large models, as an industry with a high degree of intelligence penetration, intelligent customer service is bound to become one of the main service forms in the future. According to relevant data statistics, it is expected that by 2027, the market size of China's intelligent customer service industry will reach 18.13 billion yuan, and 95% of customer service interactions will be dominated by AI technology.

Operators, Strong Competitors in the Large Model Race

The news of operators deploying large models has triggered widespread discussions:

From a timing perspective, the earliest operators to deploy were also in July last year, during a period when domestic large models were in full swing, apparently losing the first-mover advantage; entering late may be more of imitation, making it difficult to create their own characteristics... These voices seem to convey a denial of operators'