AI models developed by fine tuning LTM with view toward advanced cellular network operations
TOKYO, March 19, 2025--(BUSINESS WIRE)--SoftBank Corp. (TOKYO:9434, "SoftBank") announced that it has developed a new Large Telecom Model (LTM), a generative AI foundation for the telecom industry. The LTM is trained on diverse datasets—ranging from SoftBank’s huge network data to the design, management, and operational know-how the company has accumulated over many years. The LTM enables advanced inference in the design, management, and operation of cellular networks. Moving forward, SoftBank will further advance its research and development efforts, aiming to implement the LTM into its own operations.
SoftBank has also developed specialized AI models by fine-tuning the LTM, which is specifically designed to optimize base station configurations that enable advanced cellular network operations. The fine-tuned models were tasked with predicting configurations for actual base stations that had been excluded from the training phase, and their predictions were later verified by in-house experts to have over 90% accuracy. Compared to manual or partially automated workflows, the LTM-led approach reduces the time to make these changes from days to minutes, and with similar accuracy, indicating the potential for huge operational time and cost savings, in addition to reducing human error.
These results demonstrate that by fine-tuning the LTM for specific use cases, it will become easier to develop dedicated AI models tailored to various operational scenarios in cellular networks. The LTM also functions as a foundation for the "AI for RAN" initiative, which aims to enhance RAN (Radio Access Network) performance through AI. In the future, the LTM is expected to serve as a blueprint for network design and support the development of network optimization AI agents.
The LTM model was further optimized using NVIDIA NIM, which allows for significant performance gains for the two specialized use cases, including about a fivefold improvement in both Time to First Token (TTFT) and Tokens Per Second (TPS). Furthermore, using NVIDIA NIM provides SoftBank deployment flexibility (On-Prem and Cloud).
This technology is an implementation of the "Human AI" concept*1 envisioned by SoftBank's Research Institute of Advanced Technology (RIAT). SoftBank RIAT has proposed two approaches for utilizing AI in mobile networks, "Human AI" and "Machine AI," and has now successfully realized its vision of "Human AI". SoftBank aims to integrate various AI models developed based on the LTM with the orchestrator*2 of "AITRAS"*3, an AI-RAN integrated solution currently under development by SoftBank.
Main features of LTM
The LTM combines advanced inference capabilities leveraging large-scale data to solve network operational issues with flexible responsiveness enabled by natural language processing. Its main features are as follows:
1. Knowledge and insights as SoftBank’s mobile network specialist LTM provides knowledge and insights cultivated by SoftBank’s mobile network experts. It reflects SoftBank’s extensive network and data, along with in-depth network information annotated by in-house experts skilled in network design, management, and operation.
2. Realizing use cases specific AI models through fine-tuning By fine-tuning models based on the LTM, it is possible to develop AI models specialized for various use cases in mobile network operations. As the first implementation, SoftBank developed models specialized in generating optimal base station configurations, as described below. Its effectiveness has been verified in scenarios including generating optimal configurations for newly deployed base stations and modifying existing base station settings to accommodate sudden traffic increases expected during events.
- New base station deployment: Focusing on Tokyo, a high-density urban area, the model generates optimal configurations for new base station deployments. The model receives requests to deploy a new base station in a specific area, along with additional information such as existing base station configurations and network performance, and outputs a list of configurations recommended for the new base station.
- Existing base station reconfiguration: Assuming a special event is taking place, the model generates configuration changes for existing base stations in the surrounding area. The model receives requests to optimize configurations for a specific base station, along with additional information such as existing base station configurations and network performance, and outputs recommended configuration changes for the target base station.
3. Collaboration with NVIDIA In developing its LTM, SoftBank used the NVIDIA DGX SuperPOD for distributed training. As SoftBank moves forward towards the deployment of the LTM, SoftBank will continue collaborating with NVIDIA on NIM Microservices Optimization for Inferencing and Aerial Omniverse Digital Twin (AODT) for simulating and validating the LTM configuration changes prior to taking actions.
SoftBank will explore utilizing the LTM in its own operations, aiming to enhance mobile network efficiency, create new services, and deliver higher-quality network experiences. SoftBank will also continue to advance its research and development efforts and strengthen collaborations with partners both in Japan and abroad, thereby contributing to the further evolution of next-generation networks. In particular, the SoftBank RIAT Silicon Valley Office, which led the development of LTM in collaboration with the Japan team, will continue to grow and develop its portfolio in the USA.
Ryuji Wakikawa, Vice President, Head of the Research Institute of Advanced Technology at SoftBank said: "SoftBank's AI platform model, the 'Large Telecom Model' (LTM), developed for telecommunications operators, significantly transforms the processes of designing, constructing, and operating communication networks. By fine-tuning LTM, it’s possible to build AI models specialized for various processes and deploy them as agents. This not only optimizes and automates operational tasks but also enhances network performance through the tuning of wireless devices. SoftBank will continue to leverage cutting-edge AI technologies, aiming to deliver unprecedented levels of high-quality communication services to customers."
Chris Penrose, Vice President of Telecoms at NVIDIA, said: "Large Telecom Models are the foundation for simplifying and speeding up network operations by enabling the creation of network AI agents for specialized tasks such as network planning, network configuration and network optimization. SoftBank’s rapid innovation in developing its new LTM, leveraging NVIDIA AI technologies, sets a powerful example for telecom operators globally to redefine their network operations processes with AI."
About SoftBank Corp.
Guided by the SoftBank Group’s corporate philosophy, "Information Revolution – Happiness for everyone," SoftBank Corp. (TOKYO: 9434) operates telecommunications and IT businesses in Japan and globally. Building on its strong business foundation, SoftBank Corp. is expanding into non-telecom fields in line with its "Beyond Carrier" growth strategy while further growing its telecom business by harnessing the power of 5G/6G, IoT, Digital Twin and Non-Terrestrial Network (NTN) solutions, including High Altitude Platform Station (HAPS)-based stratospheric telecommunications. While constructing AI data centers and developing homegrown LLMs specialized for the Japanese language with 1 trillion parameters, SoftBank is integrating AI with radio access networks (AI-RAN) with the aim of becoming a provider of next-generation social infrastructure. To learn more, please visit https://www.softbank.jp/en/
