NTT’s tsuzumi 2

Large language models are getting bigger, with more parameters, computational resources and training data. But businesses don’t need massive scale – they need efficiency. That is why enterprises are looking for artificial intelligence (AI) models that can deliver high performance without adding complexity to their infrastructure.

NTT’s tsuzumi was developed to meet that need.

It is a small LLM that functions as a specialized, customizable node within NTT’s broader AI Constellation, a multi-agent, debate-style AI system. tsuzumi combines high performance with low power consumption and compact size, while also delivering world-class Japanese processing performance. It also gives enterprises the flexibility they need for domain-specific applications.

NTT is focused on resolving societal issues through the collective intelligence of smaller, efficient AIs, and we developed tsuzumi entirely in-house, in line with our AI Charter – six basic principles that inform our ethical, responsible approach to AI.

And now we have released tsuzumi 2, delivering the next generation of safe, secure, trustworthy AI, built for the enterprise. Boosting performance with seamless collaboration among humans, systems and AI, tsuzumi 2 has made major advances in contextual and semantic understanding. Customers such as Tokyo Online University and FUJIFILM are already using it.

This isn’t just a generic public chatbox; tsuzumi 2 is a high-performance, high-security and low-cost small LLM, designed for enterprise and industry use. It supports corporate digital transformation with a model that offers adapter style tuning, or lightweight fine-tuning, so organizations can adapt it to domain- or company-specific tasks.

The use of AI in today’s business environment

AI excels when data is centralized. Yet in business, information often sits in departmental or system silos, or in people’s heads, making consolidation difficult. As a result, organizations underuse AI’s potential.

tsuzumi 2’s customizability and fine-tuning capabilities have the potential to change all this. Data from across an organization can be gathered and fed into tsuzumi 2, which can then be tuned for departmental and system use – performing tasks customized for each department. Integrating data from multiple internal systems also helps the business with cross-departmental planning and other tasks.

What makes tsuzumi 2 different from other AI models

tsuzumi 2 stands out from other AI models for a number of reasons. It was built strictly in-house from the ground up, not adapted from overseas models, and it uses closely controlled training data to help ensure its trustworthiness. tsuzumi 2 supports deployment on-premises or in private clouds, which helps to keep sensitive data secure.

It performs inference on a single GPU, dramatically reducing energy use and costs. These individual GPUs can then be connected via NTT’s high-speed, low-latency IOWN network – and NTT is unique in being able to offer both AI and photonics.

In addition, it is optimized for finance, healthcare, and public sector use cases through fine-tuning and retrieval-augmented generation (RAG), and it empowers companies to develop domain-specific AI tools quickly and securely. It’s been proven in internal NTT trials to meet or surpass performance benchmarks from leading international models.

While tsuzumi was developed in Japan and is currently optimized for Japanese, it is being developed as a global solution.

How organizations are using tsuzumi 2

This isn’t just theoretical. It’s in use already – for instance, Tokyo Online University has adopted tsuzumi 2 for AI that protects student data with secure applications. The university is using it in both educational and administrative domains, for use cases including: enhancing course Q&As for students; supporting the creation of teaching materials and exams; and developing personalized guidance for course registration and career counseling. Because all data stays within the university’s network, it ensures total confidentiality.

In another example, NTT DOCOMO BUSINESS and FUJIFILM are collaborating on an enterprise knowledge management initiative that combines tsuzumi 2’s generative AI with FUJIFILM’s REiLI technology. It transforms unstructured corporate data such as contracts and proposals into structured, actionable insights. The goal of this work is to enable enhanced productivity and smarter business decision-making through secure, generative intelligence.

Responsible AI for a connected future

As we look to the future, NTT is advancing AI through our AI Constellation concept, a next-generation AI framework in which multiple, small, specialized AI models and LLMs like tsuzumi 2 are linked together to collaborate, debate, and solve complex problems – co-creating with humans involved in the multi-agent debates. NTT believes in the concept of ‘self as we,’ in which humans, nature and now technology are all connected, and one of our key goals with tsuzumi 2 is to create an AI that can naturally coexist with humans.

With tsuzumi 2 we are giving CEOs and enterprise leaders an AI model that is secure, efficient, trustworthy and sustainable. As the technology continues to develop, we envision future applications for tsuzumi 2 that include cybersecurity, intelligent debate systems, and autonomous decision support. At NTT we are continuing to advance our commitment to building an AI ecosystem based on trust, transparency and local innovation – and providing enterprises with the efficient, flexible, high-performance solutions they need to succeed.

• LLM
• Research & Development (R&D)