ZTE’s Nebula Telecom Large Model: Driving Networks Toward High-Level Autonomy

Today’s communication networks face two major challenges: rapidly increasing complexity and the growing demand for intelligence. With 5G-A commercialization and the advent of 6G, ubiquitous connectivity is driving emerging industries such as XR and Industrial Internet, accelerating business value creation. Meanwhile, the integration of new technologies such as AI large models is reshaping networks into a new intelligent paradigm.

ZTE's AIR Net high-level autonomous network solution provides full-scenario, closed-loop automation across multiple services and domains throughout the entire lifecycle with an open, decoupled design. By leveraging AI large models, intelligent agents, and digital twins, it accelerates the path to high-level autonomous intelligence, guided by three key values:

• Commercial effectiveness: Enabling end-to-end closed loops in high-value autonomous network scenarios, reducing labor, costs, and time, while supporting machine decision-making.
• O&M efficiency: Shifting from manual-driven to data-driven intelligence, boosting fault-handling efficiency severalfold.
• Ecosystem value: Driving collaboration between vendors and operators in vertical industries to meet diverse demands, creating a new "network as a service (NaaS)" ecosystem.

Building a Flexible, Dynamic, and Rigorous Intelligent Foundation

The core goal of ZTE's AIR Net autonomous network architecture is to enhance autonomous operation and optimize the closed-loop efficiency of business responses, eliminating process gaps and bottlenecks caused by manual intervention. Current networks, while capable of basic automation, such as alarm filtering and simple work order dispatch, still relies largely on predefined rules and local AI models. This approach has clear limitations in complex scenarios such as cross-domain fault localization and quality degradation analysis. To truly leap forward, the network requires an "intelligent hub" with deep network cognition—an intelligent decision-making engine powered by large models. By tightly integrating general intelligence with domain-specific knowledge, the network will transition from automation to intelligence.    

ZTE’s Nebula Telecom Large Model (see Fig. 1) addresses this with breakthroughs in four dimensions—foundation flexibility, task autonomy, knowledge dynamics, and reasoning rigor—enabling the network to evolve from rule-based automation to cognitive automation.

Full-Stack Large Model System: Seamless Engine Switching for "Foundation Flexibility"

The complexity of network environments lies in hardware heterogeneity, platform diversity, and long-tail scenarios. Through a layered, decoupled architecture, the Nebula Telecom Large Model realizes seamless engine switching, allowing the foundation model to be replaced without modifying business logic.

Leveraging its self-developed Nebula Telecom Large Model, ZTE has built a multi-agent collaboration system trained on massive, high-quality communication-specific corpora to precisely address complex network O&M challenges. With strong decoupling capabilities, ZTE’s large model solution can flexibly adapt to multiple models. At the model foundation level, it not only supports ZTE's Nebula Telecom Large Model but is also compatible with outstanding open-source models in the industry such as DeepSeek. The core advantage of this solution is its seamless engine switching, allowing users to switch model engines across scenarios without impacting performance or compatibility.

This open architecture not only avoids the risk of technical path lock-in but also forms a healthy ecosystem of "foundation model competition and selection." Operators can flexibly choose foundation models according to business needs, utilizing both the generalization strengths of general-purpose models and the precision advantages of vertical domain models, ultimately realizing the vision of "the best model for the best scenario."

Multi-Agent Collaboration: Goal-Oriented "Task Autonomy"

Intelligent agents are key to fully unlocking the capabilities of large models, solving complex problems in communication networks independently and quickly through collaboration. As network O&M transforms toward large-model-driven operations, the paradigm is shifting from "human + machine" to "machine + human". Agents interact via the language programming interface (LPI), moving beyond the traditional API-based approach of addressing fixed scenarios, thereby improving their generalization capabilities.

Today’s agents mainly complete tasks through dialogue and limited tool calls within fixed task flows. To enhance task autonomy, they need to support more advanced long-horizon autonomous planning and broader domain tool calling. Therefore, ZTE is integrating Manus-like capabilities into foundation models while continuously developing domain tools for networks to better cope with complex tasks.

ZTE is actively building an agent-centric, iterative O&M system, where multiple atomic agents—both cross-domain and single-domain—are orchestrated based on scenarios to enable global agent collaboration. At the cross-domain layer, a multi-agent collaboration center orchestrates and flexibly invokes monitoring agents across and within domains. To ensure that agents can access the required capabilities when needed, atomic capabilities are developed in a composable manner. Specialized agents then chain these atomic capabilities according to scenarios, completing tasks through cross- and single-domain collaboration.

Knowledge Graph: Dynamic Integration for "Knowledge Dynamics"

With the growing scale and complexity of 5G networks, network fault monitoring and localization face severe challenges. Multi-source heterogeneous data (e.g. logs, alarms, topology, and device configurations) are scattered across isolated systems, leading to inconsistent semantics, significant format differences, delayed updates, and serious information silos. Traditional O&M methods, which rely on manual experience and static rule bases, cannot efficiently process massive unstructured data (e.g., work order records, maintenance logs) or to perform real-time root cause analysis in dynamic network environments. Addressing these challenges requires integrating multi-source heterogeneous data and adopting more effective representation methods.

Knowledge graphs provide a way to store and represent the relationships across such data and associated knowledge, consolidating scattered information—such as O&M experience, instruction manuals, and work orders—into a unified framework. They use graph search and enhanced large model reasoning to improve the efficiency and accuracy of fault localization. Both structured data and empirical corpus data in the communications field can be stored graphically, helping large models achieve more efficient utilization and accurate reasoning.

As the "memory hub" for telecom large models, knowledge graph faces two bottlenecks: data timeliness (due to dynamic changes in network configurations) and knowledge completeness (due to variations across vendors’ devices). Current knowledge graph technologies can partially structure knowledge but still heavily rely on manual annotation and predefined ontologies, This results in low automation, poor generalization, and lagging updates, failing to meet the real-time analysis and decision-making needs of fault propagation paths in complex network environments.

ZTE proposes an intelligent O&M framework that integrates multi-source heterogeneous data with LLMs, adopting a hybrid mode of "rule constraints + LLM reasoning" to build a dynamic knowledge graph. Unlike traditional methods relying on structured data, this approach leverages LLMs to extract implicit relationships from unstructured logs. It realizes real-time knowledge graph updates and dynamic generation of fault propagation trees through a stream processing engine. This provides both theoretical support and technical guarantee for accurate fault localization  and root cause analysis in wireless networks, offering significant value in improving network reliability and O&M efficiency.

In-Depth Reasoning: Knowledge Enhancement for "Rigorous Inference"

The complexity of communication networks lies in long causal chains (from user perception to wireless access, transmission, and the core network) and multi-objective constraints (e.g., delay, energy consumption, and cost). General-purpose LLMs are prone to hallucinations and require the injection of domain-specific logical constraints through knowledge graphs. To improve inference accuracy, segmentation and  reflection can be applied during large-model reasoning.

Phased reasoning consists of three sub-stages: hypothesis generation, knowledge verification, and iterative correction. In the first stage, the large model uses multimodal observation data—such as alarm logs, performance indicators, and topology status—together

with pre-trained domain knowledge and pattern-recognition capabilities to generate an initial set of hypotheses. In the second stage, the hypothesis queue is input into the dynamic knowledge graph system for domain logical constraint verification. In the final stage, a feedback reinforcement mechanism is established based on the verification results.

With an innovative model-adaptive, difficulty-graded distillation technology, ZTE's Nebula Telecom Large Model generates chain-of-thought (CoT) corpora that are manually proofread to enable the cold start of the foundation model. This stage endows the model with complete reasoning output capabilities. High-quality question-validator corpora are then used, along with reasoning-oriented reinforcement learning algorithms, to further enhance its performance in specific complex domains.

Harnessing the Large Model for Autonomous and Accurate Problem-Solving in Complex Scenarios

To advance toward high-level autonomy, the challenges of handling long-sequence, multi-dimensional, and structured data have been a central focus of technical research. Through refining the four basic capabilities, we enable copilots and agents for various scenarios to tackle these challenges (see Fig. 2).

• Automated Closed Loop for Long-Process Scenarios

The core of an autonomous network lies in integrating technologies such as large models, intelligent agents, and digital twins, replacing manual judgment and operations with system-driven decision-making and execution to realize automated closed loops for end-to-end scenarios. Higher levels of autonomy place higher demands on these loops—shifting from a "best-effort" mode with fixed rules and manual fallback to addressing all long-tail scenarios and achieving comprehensive automated closed loops. This requires the O&M system to have generalization abilities, as well as enhanced autonomous process planning and problem-solving capabilities.

Taking cross-domain fault handling scenarios as an example, the complete process includes four stages: alarm detection, fault demarcation and localization, scheme execution, and effect verification. The traditional method requires manual data transfer and coordination across multiple systems, creating the risk of response delays. Intelligent transformation should eliminate these breakpoints and blockages, utilizing multi-agent collaboration to remove the need for manual operations. Through the flexible switching of foundation models and continuous iterations of multi-agent collaboration, the Nebula Telecom Large Model demonstrates Manus-like long-range planning capabilities and can enhance deep reasoning capabilities as industry models evolve.

• Intelligent Multi-Dimensional Decision-Making

Multi-dimensional decision-making is a key intelligent capability of communication networks. It requires balancing multiple target parameters such as bandwidth, delay, energy consumption, security, and cost in dynamic environments, and achieving globally optimal solutions through cross-layer collaboration and real-time computing. Traditional rule-based methods or single-objective optimization can no longer handle the nonlinear coupling problems in complex scenarios such as 5G/6G network slicing and edge computing. The introduction of model-based agents are reshaping the paradigm of network optimization by building a closed-loop cognitive system of "perception–inference–decision–verification."

Taking network optimization as an example, it is necessary to comprehensively consider multi-dimensional data (user distribution, traffic characteristics, and device status), establish a multi-objective optimization model, and generate optimal schemes under constraints such as delay, energy consumption, and cost.  ZTE’s Nebula Telecom Large Model integrates knowledge graphs, structured data, and in-depth reasoning to obtain optimal solutions.

• In-Depth Value Mining of Structured Data

With massive network operation data, it is necessary to break through the limitations of traditional threshold-based alarms and identify hidden fault propagation chains by constructing alarm correlation models. At the same time, historical work order data can be mined to extract high-frequency solutions and predict potential network risks.

With the support of generative AI, vertical-domain corpora are evolving from auxiliary training sets into core knowledge carriers. Their role now goes beyond traditional feature engineering, serving as a “genetic blueprint” for domain cognition. ZTE’s Nebula Telecom Large Model overcomes challenges such as multi-source heterogeneous data fusion, professional knowledge distillation, lightweight reasoning, and graph search, ultimately enabling deep internalization and rapid retrieval of telecom knowledge through a systematic methodology.

Summary and Future Prospects

The problem-solving, reasoning, and tool-invocation capabilities of telecom large mode-based agents are constantly improving. As accuracy meets production-level needs, the vision of "machine decision-making replacing manual decision-making" is gradually becoming a reality. This progress is grounded in the large model's rationality, complete logical reasoning, and ability to autonomously bridge process blockages and breakpoints. Therefore, the core task of domesticating large models lies in enhancing chain-of-thought reasoning within the communications domain, enabling precise multi-agent collaboration and tool use, and continuously expanding problem-solving capacity in real-world production scenarios. ZTE's Nebula Telecom Large Model is enhancing its capabilities in seamless engine switching, dynamic knowledge integration, and in-depth reasoning, extending its reach to complex scenarios and driving the network toward high levels of autonomy.

Looking ahead, with the commercialization of 5G-A and the arrival of 6G, future 6G networks will become AI-native, where AI applications will drive breakthroughs in intelligent perception, modeling, O&M, and resource scheduling. At the same time, network architectures, such as the data and control planes, will evolve to accelerate comprehensive AI deployment, realizing the mutual advancement of AI and 6G.