#AI-RAN #LargeTelecomModel #AITRAS
May 11, 2025
SoftBank Corp.
Blogs
In recent years, the rapid advancement of generative AI has led to the widespread adoption of large language models (LLMs) across various industries. However, the knowledge possessed by general-purpose LLMs remains broad and non-specialized, making them less effective in domains that require highly specialized and practical expertise—such as telecommunications network operations.
To address this challenge, SoftBank's Research Institute of Advanced Technology has developed the Large Telecom Model, with the goal of transferring domain-specific knowledge and real-world operational know-how in the telecommunications industry to AI. Large Telecom Model builds upon general-purpose LLMs and is further trained with SoftBank's vast and sophisticated operational data, domain expertise, standardization documents, and simulation results—making it a generative AI model specifically optimized for the telecom sector.
Press Release:SoftBank Corp. Develops a Foundational Large Telecom Model (March 19, 2025)
Specifically, a wide range of data types are integrated and utilized for fine-tuning the model, including:
● Base station configuration data (e.g., location, orientation, frequency, and various operational parameters)
● Performance data (e.g., number of connected users, throughput, signal strength, etc.)
● Public domain-specific data from sources such as 3GPP, ETSI, IETF, ITU, ORAN, and arXiv
● Simulation data generated through low-layer signal simulations
By comprehensively learning from this diverse data, Large Telecom Model evolves beyond a simple interface—becoming an AI model that can think, propose, and improve as an expert in telecommunications operations.
The most defining feature of Large Telecom Model is its design as a “Human-AI” system that complements or even substitutes for human knowledge and expertise in telecommunications operations. While conventional AI models are capable of interacting through natural language, they often lack the internalization of specialized knowledge and field experience, limiting their ability to truly take over operational tasks.
In contrast, Large Telecom Model is trained to understand domain-specific data, enabling it to perform complex operational decisions, configurations, and tuning. To draw an analogy: traditional AI is like a smart newcomer with no practical experience, whereas Large Telecom Model is more akin to a seasoned professional with deep field expertise.
Large Telecom Model can proactively propose RAN parameter settings (e.g., cell selection thresholds) based on area characteristics and traffic conditions. It supports automated configuration generation and tuning tasks.
By analyzing alert logs, Large Telecom Model can automatically identify the root causes of anomalies and suggest appropriate countermeasures. This enables real-time and accurate decision support—tasks that traditionally required significant time and expertise from experienced engineers.
Large Telecom Model can learn from customer feedback, traffic trends, and the effectiveness of past promotional campaigns, enabling end-to-end execution of targeted marketing strategies and campaign planning.
Through these applications, Large Telecom Model delivers a wide range of benefits, including workforce reduction, faster task execution, lower operational expenditures (OPEX), and more consistent decision-making. Furthermore, by encapsulating previously siloed, person-dependent tasks within the model, Large Telecom Model holds significant value from a knowledge management perspective as well.
Large Telecom Model has been evaluated across multiple use cases, with particularly promising results observed in the following two scenarios:
To handle traffic surges during large-scale events, Large Telecom Model automatically proposes configuration changes for base stations. For example, during sporting events or concerts held in stadiums, communication loads often concentrate in specific areas.
Large Telecom Model analyzes these traffic patterns and outputs both the parameters to be optimized (e.g., cell boundary adjustments) and their recommended values. In experiments, it was confirmed that Large Telecom Model could generate deployable configurations with a high accuracy of 94%, compared to the optimized data used in SoftBank's live network.
Simulation Movie
Another key use case is addressing newly developed areas where signal coverage may be insufficient, such as in urban development projects. Large Telecom Model can automatically generate suitable locations for new base stations as well as their initial configuration parameters.
Based on this approach, Large Telecom Model was also able to propose configurations for new base stations with an accuracy of 91%, significantly reducing the design and verification effort traditionally required.
These results demonstrate that the model goes beyond simple language generation, serving as a powerful decision-support tool for complex tasks in network design.
The application of Large Telecom Model is expected to expand through the following three phases:
This phase focuses on verifying the technical feasibility and business value of Large Telecom Model across representative use cases in telecommunications operations.
The model will be scaled by increasing the volume of training data and the number of parameters, enabling it to handle a broader range of scenarios with greater generalizability.
Full-scale implementation within internal operations will be pursued, with future integration into the AI-RAN unified solution “AITRAS” also under consideration.
SoftBank aims to realize a new vision of infrastructure through Large Telecom Model—one shaped by the fusion of AI and telecommunications. This vision goes beyond simple efficiency or automation; it envisions a future where humans and AI collaborate to build the next generation of network operations.
The Large Telecom Model represents SoftBank’s initiative to embed its deep operational expertise into AI, driving a fundamental transformation in how networks are managed. Rather than pursuing general-purpose intelligence, Large Telecom Model marks a step toward domain-specialized AI, bringing new forms of intelligence directly to the frontlines of telecommunications.
Going forward, SoftBank’s Research Institute of Advanced Technology will continue to advance research and development aimed at evolving social infrastructure through the integration of AI and network technologies.