SoftBank Corp.'s Large Telecom Model—
a Generative AI Foundation for the Telecom Industry—
Evolved into a Domestic AI Model
and Launched for Internal Use

Realizing a Japan-developed AI model with the homegrown LLM Sarashina

Starting in 2024, SoftBank began developing the LTM to enhance the efficiency and sophistication of mobile network operations. By leveraging the LTM, network operations that were previously manual or only partially automated can now be executed within minutes instead of several days, while maintaining the same level of precision. This advancement not only reduces human error and dependence on individual expertise; it also significantly shortens overall operation time.

Previously, LTMs were built on general-purpose LLMs, which presented challenges related to licensing and compliance. In addition, the LTMs, primarily trained on English data, struggled to accurately understand and express the complex nuances and contextual meanings of the Japanese language.

To address these challenges, SoftBank integrated Sarashina, a homegrown LLM with strong Japanese language capabilities developed by SB Intuitions, into the LTM. This integration enables a fully domestic AI process—from data training to operation—realizing a Japan-developed AI model capable of securely handling sensitive information. The new LTM can accurately understand complex Japan-specific contexts and technical terminology, providing advanced support for operational tasks and contributing to improved efficiency and convenience.

In addition, since the system can be operated within a highly secure environment, SoftBank is also considering expanding the service to external partners in the future.

Enhancing the efficiency and sophistication of network operations through LTM

SoftBank began internally using the LTM evolved into a Japan-developed AI model and is pursuing initiatives to further advance its capabilities and expand its range of applications. Verification tests are underway in use cases focused on optimizing base station configurations and improving communication quality.

For example, optimizing base station settings during large-scale events requires the accurate forecasting of rapidly changing network conditions and an understanding of the event-day environment in advance. To meet these requirements, SoftBank fine-tuned the LTM and developed its "Communication Quality Prediction Model," a specialized model designed to predict communication quality of base stations at any given point in time. By inputting parameters such as the expected number of device connections in the target area, base station frequencies, and adjustable configuration parameters, the model generates predictions of communication quality.

As a practical use case, SoftBank applied the model to predict communication quality during the Kita City Fireworks Festival held in Tokyo's Kita Ward on September 27, 2025. As this area tends to experience concentrated traffic and overlapping signals from multiple base stations, LTM was provided with more than 1,000 simulated base station conditions to forecast changes in connection volumes for each station. When the predictions were compared with actual measurements from the event day, the model was found to achieve over 90% accuracy in predicting communication quality for each frequency band.

Enhancing the LTM with reinforcement learning

To further advance the capabilities of the LTM, SoftBank upgraded it to a reasoning model that incorporates insights from Chain of Thought (CoT)^*3 generation and reinforcement learning (RL). The new Reasoning Model enables advanced inference—such as mathematical computation and logical analysis—that conventional non-reasoning models struggled to perform. In addition, by visualizing the model's thought process, human operators can now review and analyze the reasoning behind LTM's decisions.

SoftBank is also exploring new use cases for LTM that apply reinforcement learning techniques used in reasoning model development. One such initiative involves setting reinforcement learning rewards based on results obtained from radio wave propagation simulators. This creates an environment in which LTM can continuously learn to achieve optimal radio conditions. As a result, rather than simply mimicking expert judgments through traditional supervised learning^*4, the LTM is expected to surpass human expertise and achieve even higher levels of operational optimization.

SoftBank will continue to expand the application areas of the LTM while utilizing the secure training and inference environment enabled by its Japan-developed AI model that operates entirely within Japan. Through these efforts, SoftBank aims to eliminate dependence on individual expertise, reduce operational workloads, and improve operational efficiency in network management. By further promoting internal use of the LTM, SoftBank also seeks to provide higher-quality mobile network services. In addition, it will strengthen collaboration with universities and research institutions to adopt the latest technologies and enhance the accuracy and performance of its AI models.

Hideyuki Tsukuda, Executive Vice President & CTO of SoftBank Corp., said:
"The LTM is a proprietary generative AI foundation model developed by SoftBank to enhance the efficiency and sophistication of network operations. While we have long pursued automation, the LTM enables AI-driven autonomous operations that reproduce the expertise and decision-making of experienced engineers, advancing base station optimization and improving communication quality.
Going forward, with the evolution of the LTM into a Japan-developed AI model, we will securely utilize in-house expertise and sensitive data to further strengthen the reliability of our network operations."