NTT's LLM tsuzumi 2 Updated
~Achieving World-Class Performance in Processing Japanese Business Documents with Charts and Tables in a Single-GPU Environment~

News Highlights:

Background

In recent years, large language models (LLMs), including OpenAI's ChatGPT, have attracted significant attention. However, conventional LLMs require enormous amounts of training data and computational resources, creating challenges such as increased power consumption, higher operational costs, and security risks associated with handling sensitive information.

To address these challenges, NTT announced tsuzumi 2 in 2025, an LLM developed by NTT that combines a compact design with world-class Japanese language processing performance, promoting the adoption of AI among Japanese enterprises and government organizations. Leveraging its lightweight architecture, the model has primarily been deployed in on-premises and private cloud environments for handling sensitive information.

Development Results

This update focuses particularly on documents frequently used in business settings. Because sensitive information is often included in charts and tables within business documents, processing text alone is insufficient, making it necessary to process documents as images. Leveraging tsuzumi's inherently strong Japanese language processing capabilities, the update enhances comprehension of Japanese business documents containing charts and tables (Figure 1). The enhanced model further improves operational efficiency in handling sensitive information, including extracting key information from charts and tables and converting it into structured database formats.

Figure 1. Comparison of document image comprehension performance with other models

Figure 2. Example 1: Extraction of required information from forms for credit screening and other operational support

Figure 3. Example 2: Understanding cause analysis flowcharts for technical inquiry support

In addition to improvements in chart and table comprehension, the update also enhances other performance areas. Compared with other models, tsuzumi continues to demonstrate strong Japanese language processing performance (Figure 4). The update strengthens logical reasoning capabilities required for practical applications, including numerical processing. For example, the model can now perform more advanced understanding and calculations involving numerical data such as sales figures, as well as interpretation of functions included in technical documents such as API documentation. These enhancements enable support for more complex and diverse requirements.

Figure 4. Comparison of Japanese language processing performance with other models

Future Plans

The updated tsuzumi 2 will be gradually made available through companies across the NTT Group. Continued research and development of tsuzumi will support the delivery of valuable services to customers.

Related Previous News Release

About NTT

NTT is a leading global technology innovator, providing a broad range of services to both consumers and businesses. As a mobile operator and provider of infrastructure, networks, and services, NTT is dedicated to promoting a sustainable future through cutting-edge innovations. Our portfolio includes business consulting, AI-powered solutions, application services, global networks, cybersecurity, data center and edge computing, all supported by our deep global industry expertise. Generating over $90 billion in revenue and employing 340,000 professionals, we allocate 30% of our annual profits to fundamental research and development. With operations spanning more than 70 countries and regions, our clients include over 75% of Fortune Global 100 companies, alongside thousands of enterprises, government organizations, and millions of consumers.