New Intelligence Era Report
Editor: Editorial Department
New Intelligence Era Overview: A state-owned communication giant once again takes the stage with a powerful performance in the realm of large-scale models! At the recently concluded China Mobile Artificial Intelligence Ecological Conference main forum, China Mobile unveiled its independently developed "Nine Heavens" multi-modal base large model, instantly becoming the most trustworthy and industry-savvy full-stack large model in the country.
In the era of general large models, China Mobile stands at the forefront, never falling behind!
On May 24th, at the 7th Digital China Construction Summit, China Mobile Artificial Intelligence Ecological Conference themed "AI Empowerment, Building the Foundation of Wisdom for the Nation," China Mobile introduced its latest self-developed "Nine Heavens" multi-billion multi-modal base large model.
With independent research and industry expertise, the "Nine Days" base large model can be considered the most reliable and industry-savvy full-stack large model in the country.
Unlike other companies, central enterprises develop large models for key industries in the national economy, aiming to empower the country's infrastructure transformation and "inject wisdom into industries, build a foundation for the country."
China Mobile truly deserves the title of "leading central enterprise in artificial intelligence."
During the AI Dialogue Conference, when faced with difficult problems, AI Chat sought human help independently.
On the day of GPT-4o's release, it showcased to everyone the powerful capabilities of AI voice assistants, bringing Samantha from the movie "Her" into reality.
And the live demonstration at China Mobile's event was just as impressive as OpenAI.
Nine Days · Nine Nine
First up was "Nine Days · Nine Nine," a customized personal assistant.
The two presenters showed everyone present how a real AI assistant can significantly boost efficiency for its users.
On one side was Xiaotian, on the other was Jiujiu.
The young man asked his assistant "Xiaotian," "I'm a bit nervous, what should I do?"
The caring Xiaotian promptly came to the young man's aid, cheering him on.
The assistant "Jiǔ Jiǔ" next to the girl helped her generate "3 junior high school math questions about quadratic functions," and provided detailed solution steps upon further inquiry, effortlessly guiding the junior high school student.
Next, Jiǔ Jiǔ also created an image for Children's Day on June 1st, with a background of "Nine Sky Blue".
Finally, Jiu Jiu also completed a very practical function: telling it its own schedule, it can help arrange the schedule, such as adding a meeting at 9 a.m. the day after tomorrow.
Even more impressive, the two AI assistants, Xiaotian and Jiujiu, directly connected on the spot to discuss meeting requirements, completely independent of their owners.
Jiutian·Shanxue
Another participant in the presentation was "Jiutian·Shanxue."
As a product based on knowledge data and expert guidance, it can assist us in professional industry analysis or business insights.
Moreover, if we have a task that needs improvement, it can drive conversations between the model and experts for corrections.
For example, if it encounters a network operations issue while drafting a proposal, it will actively seek input from human experts.
After receiving advice from two experts, it can automatically proceed with the relevant updates and modifications.
Nine Days Network
"Nine Days Network" shows us how complex systems achieve global intelligent perception.
During the demonstration, the venues of the Hangzhou Asian Games and surrounding areas were replicated in a 3D twin digital world, with a visual rendering of the signal coverage of the communication network.
Based on this, "Nine Sky Network" can perceive network status, device status, and user experience.
The overall monitoring and perception results can also be presented visually. The signal halo makes the network status clear and visible everywhere. In the image below, the yellow area indicates good signal, while the red area shows signal congestion. It can even track a specific person on the road, analyzing from the perspective of individual users.
When inspecting equipment, maintenance personnel can wear AR glasses, which enhance their visual monitoring capabilities to assist in analyzing the equipment's condition. The analysis results can be directly overlaid within the maintenance personnel's field of vision.
In addition, remote experts can also be involved to provide guidance on the inspection and maintenance of the equipment.
The showcased products and capabilities above all stem from the support of China Mobile's "Jiutian" trillion-parameter multi-modal base model.
To achieve such a complex and intelligent base model, continuous technological breakthroughs are required in both model architecture and parameter scale.
Since 2013, China Mobile has been engaged in independent research and development of artificial intelligence, delving deep into the same field for a decade, accumulating continuously, enabling the "Jiutian" large model to take a leading technological position.
The "Nine Skies" base large model adopts a structure that combines pure decoding with multiple experts, innovating the model's learning approach - by modeling structured data and conducting progressive learning.
The number of parameters in the model has continuously increased from billions to the latest trillion level.
To adapt to different application scenarios, "Nine Skies" provides model versions with different parameter sizes such as 1.5B, 3B, 7B, 13.9B, 100+B, which can be flexibly deployed on various devices and platforms.
Moreover, China Mobile is also brewing a larger training plan for the next step, increasing the dataset scale from 5 trillion to 20 trillion tokens to further enhance the model's capabilities.
With its cutting-edge architecture and vast parameter count, "Nine Skies" has shown outstanding performance in international competitions:
In the field of speech synthesis, it achieved first place internationally in the 2024 INTERSPEECH Speech Synthesis Vocoder track.
In terms of visual evaluation, it also excels, winning first place in visual understanding in the 2023 CVPR VSS (Visual Semantic Segmentation) track.
Compared to the MME-P rankings, the generation capability of the "Nine Heavens" base large model is equivalent to the level of the 9th globally.
In addition to impressive competition results, the evaluation outcomes in various tests are also outstanding.
In well-known multilingual understanding MMLU, encoding HumanEval, and mathematical GSM8K benchmarks, the "Nine Heavens" base model has achieved commendable performance.
Furthermore, on the Chinese C-Eval, MBPP, and BBH benchmarks, the base large model has gained a leading edge.
In addition, in 11 industry sectors, China Mobile conducted 124 subcategories on the pedestal model, testing over 80,000 exam questions for university or professional qualifications exams, with results shown in the figure below.
The "Nine Heavens" pedestal model achieved the highest scores in all 11 sectors.
The "intelligent" performance at these conferences, as well as the impressive results in competitions and evaluations, fully justify its positioning as "the most knowledgeable in the industry."
Intelligentization of Complex Systems
As a full-stack large model, "Jiutian" has oriented its capabilities towards various complex industry demands, proposing the concept of "intelligentization of complex systems" with the aim of "injecting intelligence" into industries.
Talking about the "intelligence" of large models is easy, but implementing it is quite challenging, requiring overcoming many obstacles and barriers.
In order to achieve intelligentization of complex systems, Chief Scientist Feng Junlan and other experts from China Mobile Group have put in a lot of effort.
If AI is likened to a vast ocean, current intelligence only scratches the surface. To delve into deeper waters, besides general capabilities like reasoning, generation, and inference, it also requires perception, prediction, diagnosis, control, and decision-making for complex systems.
Most complex systems are chaotic systems. How can one transition from the micro to the macro, from chaos to clarity?
This necessitates structured data. In complex systems, there are numerous devices and vast amounts of data.
Furthermore, there are multi-source data, enabling AI to see what I see, act as I act, and hear what I hear.
Effectively representing these signals and integrating them with general modal data is the key to truly achieving AI in complex systems.
In summary, the intelligentization of complex systems requires addressing two major issues: being large but unstable, and being diverse but uncoordinated.
To address this, China Mobile has employed various strategies and technologies in the development process of the "Jiutian" base large model, including multi-layer information reinforcement, traceable information field consistency verification, and collaboration between large and small models.
The combination of these technologies transforms the "Nine Heavens" model into an intelligent entity tailored for "atomized tasks."
Beyond the understanding, generation, and reasoning capabilities of cutting-edge large models, it further extends to more "human-like" abilities such as perception, prediction, diagnosis, control, and decision-making based on this foundation.
From micro data in every industry to macro ecosystems of cross-industry models, extracting clear, simple solutions from chaotic, complex problems is the goal that "Complex System Intelligence" aims to achieve.
Connecting Modal Sources
One reason why "Nine Heavens" positions itself as a universal "full-stack" large model is to bridge the gaps between different modalities, enabling intelligent reasoning among "all modalities" in complex large model systems.
It encompasses not only common modalities like speech, language, and vision but also structured data, making modal designs more tailored to various practical scenarios in different industries.
Another reason is that China Mobile leverages its leading position in the communication field, allowing large models to be flexibly deployed on various signal sources and devices.
Current large models, due to reasons such as excessive parameters and low device compatibility, generally only perform inference on servers or PCs. However, China Mobile, which can provide comprehensive cloud-edge services, challenges this limitation.
The "Nine Heavens" large model can be deployed not only on common mobile electronic devices and IoT sensors like smartphones, tablets, cars, and wearables but also on various complex signal sources, including radar, infrared, acoustic, wireless radio frequency, and more.
These complex signals, though imperceptible to humans, are crucial for some foundational industries such as industrial production, communication, and services.
The strong adaptability to various signals allows "Jiutian" to easily enter various industries and quickly apply them in real-world scenarios.
Focusing on key industries
With large-scale models emerging in domestic and international markets, why is it that only "Jiutian" dares to claim to be the "most industry-savvy"?
From model positioning, training data to AI ecosystem, and application scenarios, the research and development team of "Jiutian" has never forgotten its original intention of "focusing on industries."
Therefore, "Nine Skies" is positioned to serve complex industry systems, playing a crucial role in government governance, industrial production, livelihood services, and distinctive communication areas, providing a foundational model.
From the outset of training, the "Nine Skies" foundational model is closely linked to the industry - integrating not only professional knowledge from eight major industries such as communication, energy, construction, and transportation but also with industry data accounting for 10% of all pre-training data, surpassing all domestic and international counterparts.
In addition, when it comes to processing structured data, the focus has been on applications in industries such as marketing and travel, enhancing the practical experience of the model's network intelligence. This approach has led to a 10%-15% improvement over traditional methods, addressing the challenge of large models struggling to accurately handle structured data.
In the field of visual understanding, the training data includes 70 million road surveillance videos, as well as over 20 billion unique industry-specific data such as images, text, and videos. This enables the models not only to "see clearly" and "understand," but also to "see accurately."
Building upon the foundation of the "Jiutian" base model, China Mobile has established a comprehensive system of large models and AI ecosystem for all industries, developing a total of 15 industry-specific large models.
Comprehensive Support for Independent Ecosystem
As a key player in the AI industry, China Mobile stands at the forefront of the AI technology wave, demonstrating its determination for independent research and development through concrete actions, showcasing the responsibility expected of a state-owned enterprise.
Throughout the development of the "Nine Heavens" large-scale model system, the research team has also been committed to achieving the "domestication" of the AI ecosystem in areas such as computing power, chips, frameworks, algorithms, and more.
At the chip level, the "Nine Skies" series of large models has tackled a major training challenge—enabling smooth conversion between heterogeneous chips.
Across 17 different chips produced by 11 manufacturers, the team has completed a series of tasks including adaptation, migration, testing, and standard alignment, enabling models to smoothly transition and continue training across heterogeneous chips.
To accelerate the development of the domestic AI industry, China Mobile has also proposed the "Ten Thousand, Thousand, Hundred" Intelligent Base Plan, aiming to build a platform with "Ten Thousand Cards for High Computing Power, Thousand Billion Models, and Hundred Exchange Platforms."
On the computing power front, China Mobile will establish 3 20,000-card smart computing centers and 12 regional smart computing centers.
Currently, the super-large single-unit smart computing center built in Hohhot has surpassed all operators globally in scale, becoming one of the "Top Ten Super Projects" of central enterprises.
With the large-scale implementation of AI, the demand for inference computing power has surged. China Mobile will comprehensively upgrade the 1,500 deployed edge computing nodes, enhancing the training and inference computing resource ratio to 1:10, meaning 1 card for training and 10 cards for inference.
By the end of the year, the total network training cards will reach over 100,000 cards, with a scale exceeding 30 EFLOPS.
Next, China Mobile will also upgrade the central nervous system of the intelligent network and build a new guarantee for the Internet of Things.
On the "Baihui Platform," a rich collection of small and large models, datasets, and AI native applications will be gathered, integrating various functions such as model experience, download, tuning, deployment, application, evaluation, etc., providing entrepreneurs and developers with a "one-stop" service platform.
Model Security and Trustworthiness
The reason why the "Jiutian" series of large models are "trustworthy" lies not only in their support for domestically developed models but also in their high standards and strict requirements for model security.
To address common pain points such as "forgetting" and "illusions" faced by large models in applications, the team has applied key technologies such as knowledge elevation fusion and continuous dynamic learning during training.
Furthermore, the models are equipped with comprehensive security mechanisms for data security, system security, etc., truly achieving security, trustworthiness, and controllability.
This relentless pursuit of security has ultimately received official recognition.
The "Jiutian" natural language interaction large model has also passed the "Generative Artificial Intelligence Service Filing" and the "Domestic Deep Synthesis Service Algorithm Filing," becoming the first large model developed by a central enterprise to complete dual filings.
In tests conducted by the China Software Evaluation Center, the "Jiutian" large model also received the highest level (Grade A) of security certification.
Over the past decade, there has been a leap in development.
Behind a series of outstanding achievements lies the ambition of the entire group to establish itself as a world-class information service technology innovation company and to bravely shoulder the important task of national scientific and technological innovation. This ambition is supported by the unwavering determination of the research and development team to strive forward day by day.
Dr. Feng Junlan, an IEEE Fellow and Chief Scientist of China Mobile Group, is a leading figure in the field of AI.
She joined the China Mobile Research Institute in September 2013, at a time when the telecom operator had already sensed the future potential of AI.
Over the course of a decade, Dr. Feng Junlan led her team in the step-by-step research and development of AI technology, witnessing the company's "leapfrog" development in the field of AI.
China Mobile has established the "Jiutian" Artificial Intelligence Research Institute, introducing 2 top experts, over 150 AI professional Ph.D.s, forming a core AI team of 2000 people. They are conducting full-stack research and development in AI fundamental technologies, algorithms, platforms, and applications, with independent innovation capabilities in large model data construction, pre-training, fine-tuning, inference, and other end-to-end technologies.
Last year, China Mobile successively released vertical industry large models such as "Jiutian·Customer Service," "Jiutian·HaiSuan e-Government," "Jiutian·Network," as well as the "Jiutian·Zhongqing Base Large Model" open for collaborative development in key industries, making a strong voice in the large model arena.
In addition to the products and models released, China Mobile has also achieved fruitful research results—originally proposing systematic Artificial Intelligence (Holistic AI) technology, receiving wide recognition and support from various sectors of industry, academia, and research.
Won the first prize of the Science and Technology Progress Award of the China Institute of Communications, applied for over 930 patents, published more than 160 papers in top conferences and journals, and achieved top 5 rankings in about 20 major international competitions.
Looking ahead, as the theme of this conference "Empowering AI, Building the Nation's Foundation" suggests, China Mobile will continue to advance the mission of AI "national team," bringing the latest advancements in the field of AI to households and industries across the nation.
