Telecom Large Models: The Innovative Force Leading Future Network Intelligence

The deep integration of AI with next-generation information technologies has become a key driver of autonomous network transformation. With the widespread application of large models in the networking field, telecom large models are reshaping network operation and service innovation, driving the intelligent evolution of networks.

Proliferation of Telecom Large Models and Their Applications in the Large Model Era

Since 2023, large language models (LLMs), represented by ChatGPT, have ignited a global technological race. In early 2025, DeepSeek released its open-source model DeepSeek-R1, which rapidly gained traction across verticals such as finance, healthcare, and smart manufacturing, due to its strong reasoning capabilities and cost efficiency. In the telecom industry, the rise of new businesses—including the low-altitude economy, the Internet of Things (IoT), vehicle-to-everything (V2X), and immersive extended reality (XR)—as well as operators’ demands for digital-intelligent transformation, has accelerated the adoption of generative AI (GenAI).  This has fueled rapid growth in autonomous networks and intelligent agent applications built on telecom large models.

The rapid development of telecom large models has been driven by active investments from both operators and equipment vendors. Major Chinese operators are leveraging their data advantages to accelerate the deployment of industry-specific large models. China Mobile has released its self-developed AI platform, "JiuTian", which accumulates 450 AI capabilities. China Telecom's Xingchen large language, speech and multimodal models have completed “dual recordation” of algorithms and services. China Unicom's Yuanjing large model has developed over 40 industry-specific models. According to IDC, the market size of large models in China's telecom industry grew by 67% year-on-year in 2024, with multimodal and scientific computing models emerging as new growth points.

On the equipment vendor side, ZTE is pursuing a dual-track approach of "full-stack in-house R&D + ecosystem collaboration." Its Nebula Telecom Large Model, available in versions ranging from 7B/14B to 100B parameters, supports scenarios such as network operation and maintenance, fraud detection, and signaling analysis through an architecture of "Nebula large model + agent factory + serialized applications".

Additionally, the open-source ecosystem fosters technological democratization and advances the development of telecom large models—DeepSeek's open-source strategy has significantly lowered industry entry barriers. By Q1 2025, over 60% of telecom enterprises in China developed customized solutions based on open-source models. This "open sharing + vertical specialization" model is reshaping the innovation path of intelligent network technology.

Telecom Large Models Reshape Autonomous Network Evolution  

Large Models Reconstruct the Intelligent Engine of Autonomous Networks

Autonomous networks achieve automation and intelligence through a "three-layer, four-closed-loop" architecture (Fig. 1). To advance toward L4 and L5 autonomous networks, TM Forum has outlined key enabling technologies, including network AI large models, trustworthiness technology, and digital twins. These technologies will support the AI requirements embedded in 3GPP networks, driving autonomous networks toward higher levels of intelligence. In this evolution, telecom large models serve as the intelligent engine. Together with the data and digital twin engines, they form the digital-intelligent technology foundation of autonomous networks.

Large Models Reconstruct the Technical Architecture of Autonomous Networks

Telecom large models are redefining the technical architecture of autonomous network evolution into a three-tier system comprising "full-scenario AI paradigms + digital-intelligent engines + intelligent agents." This architecture introduces serialized models layer by layer, facilitating embedded AI within networks and full-scenario AI integration. For example, in the transition from single-domain closed loops to cross-domain collaboration, operators can begin by achieving closed loops in single-domain optimization scenarios—such as wireless network optimization—by leveraging the intelligent analysis capabilities of large models. This can then be expanded to end-to-end, cross-domain scenarios for global network management, characterized by "traceable history, visible reality, and predictable future." Enabled by the integrated "perception–decision–execution" capabilities of telecom large models, successful applications have been demonstrated in intent understanding, autonomous learning, long-process closed-loop optimization, and accuracy improvement.

For instance:

• In planning: For 5G-A network construction, digital twin tools assisted by large models can simulate over 100,000 channel combinations, shortening planning cycles.
• In operation and maintenance: China Mobile's intelligent O&M system, based on DeepSeek, has significantly improved fault localization efficiency, paving the way for minute-level fault response.
• In operations: ZTE's fraud detection large model analyzes communication behavior patterns to identify new types of telecom fraud, greatly reducing false positives.

Large Models Reconstruct the Agent Paradigm of Autonomous Networks

Telecom large models have given rise to a new generation of intelligent agents for networks, shifting traditional autonomous network O&M toward an agent-based paradigm. With capabilities in autonomous learning, task execution, and multi-task collaboration, these intelligent agents provide a new pathway for telecom operators' intelligent transformation. By deploying agents, operators can achieve highly automated network management and optimization, enhancing  self-perception and self-healing capabilities, thereby significantly reducing operational costs and improving network reliability.

The release of OpenAI's o1 and DeepSeek's R1 models—both excelling in complex reasoning tasks—has inaugurated a new paradigm of "slow-thinking" large models, offering innovative solutions for network O&M and intelligent

agents. These models, along with their reasoning-enhancement technologies, enable task orchestration, problem analysis, flexible decision-making, and optimized execution in complex business scenarios. Meanwhile, agent applications are shaping a full-scenario AI paradigm across areas like interaction methods, interface designs, product architecture, business capabilities, and development and delivery. This agent paradigm accelerates the path toward achieving L4 autonomous networks.

Telecom Large Models Evolve Toward Network World Models

The future evolution of telecom large models will advance toward stronger performance, lower resource consumption, and diversified interaction modes, while also progressing toward network world models that are deeply integrated with network services.

Innovation Directions for Telecom Large Models

Innovation in telecom large models is advancing along several key directions: model architecture innovation, multimodal fusion, domain-augmented reinforcement learning, and the further convergent evolution of models and agents.

• Model Architecture Innovation

In telecom large model scenarios that prioritize localized deployment, architectural innovation is essential to achieve higher performance and smaller model sizes. It is worth considering the introduction of a spatio-temporal decoupled routing mechanism on top of the traditional Mixture-of-Experts (MoE) architecture, integrating temporal sequence models with spatial data to construct a spatio-temporal network model, which can then be further fused with language models.

In terms of coordination mechanisms, a bidirectional knowledge distillation approach can be employed between central large models and edge small models. The central model extracts global features (e.g., nationwide network traffic patterns), while edge models capture local characteristics (e.g., regional base station deployment differences), enabling collaborative model evolution through a federated learning framework with dynamic weight exchange.

•  Multimodal Fusion

Future large models will integrate multimodal data—including text, speech, images, network signaling, alarms, and logs. This fusion will enhance the accuracy of  network state perception and enable more immersive interaction methods and personalized customer service solutions, ultimately improving the user experiences.

•  Lightweighting and Edge Deployment

Large parameter models can be compressed using quantization and distillation to support low-power inference at edge nodes, meeting the real-time requirements of industrial IoT.

• Domain-Specific Reinforcement Learning

Enhance large models' expertise in vertical scenarios by incorporating telecom-specific knowledge graphs, such as protocol stack rules and fault case libraries.

Co-Evolution of Digital Twins and Large Models

The co-evolution of large models and digital twin technology is a form of bidirectional empowerment. Digital twins provide telecom large models with virtual-physical mapping testbeds and synthetic data, supporting twin-driven model training. In turn, large models enhance the simulation and reasoning capabilities of digital twins, equipping them with dynamic optimization abilities. While digital twins enable real-time mapping of physical network states, large models can generate optimal resource scheduling strategies through reinforcement learning. This enables a  "digital twin–model training–network deployment" closed loop for network functions, transforming networks from passive monitoring to proactive design.

Toward a Unified Network World Model

In the 6G era, telecom large models will evolve into "network world models," enabling unified representation and prediction of physical networks, business demands, and user behaviors, and serving as the intelligent core of future networks (Fig. 2).

Key features include:

• Multi-domain unified modeling: Integrate wireless, core, transport, and application-layer data with spatio-temporal digital twin models for global modeling.
• Causal reasoning capabilities: Shift from correlation analysis to causal inference.
• Autonomous evolution mechanisms: Enable the self-iterating evolution of models through continuous learning and feedback loops, allowing adaptation to the dynamic demands of new services.
• Twin closed-loop: Future intelligent agent networks will possess end-to-end digital twin closed-loop capabilities in various scenarios.   

Outlook

Telecom large models are evolving from technical tools into the core engines of network intelligence, reshaping the technological foundation of autonomous networks. They are driving a shift from single-domain optimization to full-domain collaboration, and from manual intervention to autonomous decision-making.

With the advancement of 6G network intelligence architectures, the integration of network world models—combining digital twins, edge inference, and causal learning—will propel autonomous networks toward a qualitative leap, from "functional autonomy" to "cognitive intelligence."

This intelligence revolution, powered by telecom large models, is transforming the technological form of network infrastructure and redefining the value boundaries of autonomous networks.