NTT Updates tsuzumi 2 LLM: Enhanced Chart Understanding and Logical Reasoning

## NTT Updates LLM tsuzumi 2

Realizing world-class chart-inclusive Japanese business document processing performance on a 1-GPU environment.

Key Highlights:
- Update to NTT's LLM 'tsuzumi 2', which boasts world-class Japanese processing performance despite its lightweight nature.
- Leveraging the inherent strengths of 'tsuzumi' in Japanese processing to achieve high understanding capabilities for charts, graphs, and tables frequently used in Japanese business documents.

NTT Corporation (Headquarters: Chiyoda-ku, Tokyo; President and CEO: Akira Shimada; hereinafter 'NTT') is updating its NTT-version LLM 'tsuzumi 2', which prides itself on world-class Japanese processing performance while remaining lightweight. Because of its lightweight design, tsuzumi 2 has been favored for use cases handling sensitive information in on-premises/private cloud environments. This update focuses on the fact that documents containing sensitive information often include charts, and it has been developed to understand visual meanings by reading documents with charts as images. In addition to chart understanding, logical reasoning capabilities necessary for practical use, such as numerical processing, have been strengthened.

By prioritizing the understanding of Japanese business documents, the model has achieved world-class performance compared to other models in the same size range.

### 1. Background
In recent years, while attention toward Large Language Models (LLMs) like ChatGPT has been growing, conventional LLMs require massive amounts of training data and computational resources, leading to challenges such as increased power consumption, rising operational costs, and security risks in handling sensitive information. To address these issues, NTT announced the NTT-version LLM 'tsuzumi 2' in 2025, which boasts world-class Japanese processing performance despite its lightweight nature, and is promoting the adoption of AI in domestic enterprises and administrative organizations. Leveraging its lightweight features, it is primarily used in on-premises/private cloud environments for operations handling sensitive information.

### 2. Development Achievements
This update focuses particularly on documents frequently used in business settings. Business documents often contain sensitive information in chart formats, making it necessary to process documents as images rather than just processing text. Leveraging the high Japanese processing capability inherent to 'tsuzumi', we have strengthened the understanding of Japanese business documents with charts. This will further streamline operations related to handling sensitive information, such as extracting important information from charts and digitizing it into databases.

Furthermore, updates were made to capabilities beyond chart understanding. The model continues to show high Japanese processing performance in comparisons with other companies' models. This update strengthens logical thinking capabilities, including numerical processing required for practical tasks. For example, understanding and calculating numerical information such as sales figures, and interpreting functions contained in technical documents like API documents can now be performed at a higher level, making it possible to respond to more complex and diverse requests.

### 3. Future Outlook
The updated tsuzumi 2 will be provided as a service sequentially through NTT group companies. We will continue our R&D of 'tsuzumi' to provide valuable services to our customers.