When NTT introduced its home-grown large language model "tsuzumi" in late 2023, you could say that the philosophy behind it felt slightly... counter-intuitive?

When it comes to generative AI, while most of the tech world was thinking about "bigger is better," NTT looked at things differently. Instead of building a huge, all-knowing model and then trying to find ways of using it, they thought about it from the other way around.

If we're going to make an LLM for organizations with real budgets and real hardware constraints, what does it actually need to look like?

That’s how tsuzumi was born. And the recently announced tsuzumi 2 is a major upgrade.

Gigantic LLMs Have Issues

When the generative AI boom began a few years ago, NTT’s researchers felt the same excitement everyone else did, but they also foresaw some practical issues. They knew that many organizations wanted to use AI to work with their own internal documents, sensitive data, and specialized knowledge; fine. But NTT also knew that the most well-known LLMs needed a large infrastructure setup, consumed unthinkable amounts of power, and required everything to be uploaded to the cloud. Cost? High. Technical requirements? Complicated. Security implications? Not optimal.

Contrast with tsuzumi: a compact, efficient model that could be deployed close to the data, including on premises, and wouldn't demand that everything be uploaded to another location.

tsuzumi: A Practical Partner

After tsuzumi became available commercially in March 2024, there was a clear pattern of how customers wanted to use it: for practical jobs. Most users didn't need it to tell them how to live their lives, give relationship advice or answer questions about the top goalscorers for AC Milan in the 1980s. They just wanted to do things such as loading internal manuals, guidelines, reports, and know-how into the model to get summaries, make searches, and answer questions relating to the organization. In other words, users were doing the same things they'd done before the development of LLMs, but much more accurately and efficiently. That's the whole point, right?

Featured at the 2025 NTT R&D Forum held last November, tsuzumi 2 does an even better job than its predecessor. One of its biggest upgrades is the new model's ability to work with long documents and complex context.

Here's an example: the way it can understand corporate documents.

Corporate documents are written by humans who have competing priorities and don't always set their thoughts down in the most logical way. They can be layered and inconsistent, with important decisions and guidance buried deep in the text, and assumptions made with a limited amount of context. tsuzumi 2 is better than its predecessor at understanding and following that context, which means that users no longer have to spend time explaining what they want and re-prompting multiple times. It's also better at avoiding misunderstandings.

And It Specializes, Too

Need a large language model that understands you and your organization's way of working in a very specific way? tsuzumi 2 loves to specialize.

Many organizations want an AI system that understands their field and can communicate with them in their own parlance, using the same terms. NTT gets it! tsuzumi 2 was designed so that extremely specific models can be up and running with relatively small amounts of additional training data. It adds up to faster customization and lower effort.

tsuzumi is Au Fait With Cybersecurity

Here's another big strength of tsuzumi 2: security. It's small enough to be run on a single GPU. Rather than having to run everything through the cloud, entities using tsuzumi 2 can deploy it inside their own network. End result? Less data movement, less potential for exposure, and fewer cybersecurity risks.

Another thing to note about the development of tsuzumi is the data NTT used to create it. Both the original model and tsuzumi 2 were made from scratch using data that NTT owns or has the rights to use. No intellectual property was stolen! Some LLMs have been accused of just scraping the web, taking copyrighted material and dealing with the consequences later; that's not tsuzumi.

All well and good, but how does this affect me? It means that NTT has control over training data, updates, and licensing, so that users know the tools they rely on won't be caught up in tricky copyright issues. Model updates can be delivered without any legal problems, so users can be confident it will keep evolving safely over time.

Artificial Intelligence Without Friction

tsuzumi 2 is about reducing friction. Less time spent searching through documents, less rewriting of documents, and easier creation of first drafts. It doesn't pretend to be all-knowing or have godlike powers; that would be weird. Rather, it supports human staff, so they can focus on their judgment, context, and decision-making.

Imagine that. A generative AI system that fits in with your day-to-day work. A reliable tool that makes your working day a little easier.

Innovating a Sustainable Future for People and Planet

For further information, please see this link:
https://group.ntt/en/newsrelease/2025/10/20/251020a.html

If you have any questions on the content of this article, please contact:

NTT, Inc.
Public Relations
ntt-pr@ntt.com

Daniel O'Connor

Daniel O'Connor joined the NTT Group in 1999 when he began work as the Public Relations Manager of NTT Europe. While in London, he liaised with the local press, created the company's intranet site, wrote technical copy for industry magazines and managed exhibition stands from initial design to finished displays.

Later seconded to the headquarters of NTT Communications in Tokyo, he contributed to the company's first-ever winning of global telecoms awards and the digitalisation of internal company information exchange.

Since 2015 Daniel has created content for the Group's Global Leadership Institute, the One NTT Network and is currently working with NTT R&D teams to grow public understanding of the cutting-edge research undertaken by the NTT Group.