tsuzumi 2
: How to use it, features, and the business problems it solves
Add bookmark
What is tsuzumi 2?
A fully domestic Japanese large language model (LLM) built from scratch by NTT, Inc. At 30 billion parameters (30B) it runs inference on a single GPU, so it can be deployed on-premises or in a private cloud without sending confidential data outside. It is strengthened in finance, healthcare and public-sector knowledge and in RAG search, summarization and information extraction from documents, and a May 2026 update added reading of documents with charts and tables. Available since October 20, 2025, and selected for trial use in Japan's Digital Agency "Government AI". It is not a self-serve web service: NTT Group companies provide it individually to businesses and local governments as part of their solutions, and pricing is not published.
Business problems it solves
About "tsuzumi 2"
What is tsuzumi 2?
tsuzumi 2 is a fully domestic Japanese large language model (LLM) developed by NTT, Inc. (Chiyoda-ku, Tokyo). Rather than building on an overseas open model, NTT trained it from scratch, and it became available on October 20, 2025. It is the next generation of "tsuzumi", announced in 2023.
It has 30 billion parameters (30B) and is light enough to run inference on a single GPU of the 40GB class. This means it can run on a company's own servers (on-premises) or in a private cloud without sending data to an outside cloud. NTT explains that it is mainly used for work that handles highly confidential information (as of May 2026).
This is not a service that individuals can sign up for on the web. tsuzumi 2 is provided individually by NTT Group companies such as NTT West, NTT DATA and NTT DOCOMO BUSINESS, built into their solutions for businesses and local governments. Pricing is not published. If you are looking for a Japanese LLM you can use right away, see "How it differs from other Japanese LLMs and LLM services" below.
Packing Japanese business skills into a model that runs on one GPU
According to NTT's laboratories, tsuzumi 2 was designed to run on GPUs of a slightly older generation with 40GB of memory or less. When you pay per API call, costs grow as usage grows. NTT expects more cases where running the model yourself is cheaper, and more cases where AI agents are given highly confidential information, so it focused on the balance between operating cost and performance.
Source: NTT news release (October 20, 2025), material "AI For Quality Growth: Release of tsuzumi 2" (the following figure is also from NTT's announcement materials)
NTT estimates that inference with tsuzumi 2 needs one A100 40GB-class GPU (about 5 million yen), cutting inference costs to roughly 1/10 to 1/20 compared with larger models.
On performance, NTT focused on three tasks that make up most business use. According to the laboratories, question answering over documents (RAG search and summarization) and information extraction and summarization from documents account for 80% of use cases. It also improved the model's ability to follow output-format instructions (instruction following), such as summarizing to fit a report template or outputting extracted information as JSON.
How to use it
tsuzumi 2 is provided to businesses through NTT Group companies. Based on the official announcements, the adoption process looks like this.
- Consult an NTT Group company tsuzumi 2 is proposed and delivered as part of NTT Group companies' solutions. For example, NTT West says that proposals and contracts for its on-premises solution are handled through its sales offices.
- Decide where to run it Choose between on-premises (your own servers) and a private cloud. NTT West, together with Fsas Technologies, announced (on August 26, 2026) that it plans to offer, from September 2026, a solution that runs tsuzumi 2 on the on-premises generative AI platform "Private AI Platform on PRIMERGY".
- Tailor it to your business Turn it into a model for your company with RAG over internal documents and fine-tuning with additional industry and business knowledge. NTT says that on a finance-related exam, tsuzumi 2 reached the passing line with about one tenth of the additional training data another model needed.
- Build it into business applications Embed it in applications such as internal chat, document search and summarization, or extracting information from forms. NTT Group companies also handle application development and implementation support.
Features
01Reading and summarizing documents
- RAG search and summarization — Question answering and summarization that searches internal documents. NTT says it confirmed performance equal to or better than other companies' advanced models when handling inquiries about its internal financial system
- Information extraction — Pulls the required items from documents and outputs them in a specified format such as JSON
- Long-document reading — Better understanding of complex context and intent in corporate documents, and better reading of long texts than the previous version
02Reading charts and tables (May 2026 update)
- Understanding documents with figures — Reads documents as images and understands the meaning of charts, graphs and tables
- Handling numbers — Stronger logical reasoning, such as understanding and calculating figures like sales amounts and interpreting functions in API documentation
03Tailoring to industries and companies
- Domain knowledge — Trained with strengthened knowledge of finance, healthcare and the public sector (local governments)
- Specialized tuning — Makes it easier to build models specialized for a company or industry with RAG and fine-tuning
The May 19, 2026 update added support for the charts and tables common in business documents. Because confidential information is often contained in tables and forms, the model can now read documents as images.
In NTT's example, it extracts the payment amount, withholding tax amount and social insurance premiums from an image of a Japanese withholding tax slip. The intended use is picking out required items from forms in work such as credit screening.
Pricing
As of October 24, 2026, no price list for tsuzumi 2 is published on the official websites. NTT says it proposes and implements AI solutions to fit public or closed environments and each customer's business and data requirements, so the cost depends on the configuration you adopt.
For reference, NTT's announcement materials put the hardware for inference at about 5 million yen (one A100 40GB-class GPU; operating costs not included). This is not a model usage fee but an estimate of hardware costs if you prepare your own server.
Adoption and reputation
01Government
- Digital Agency "Government AI" — Submitted by NTT DATA and selected in March 2026 as one of the domestically developed LLMs for trial use in Government AI. From fiscal 2026 it will be tested in the generative AI environment "Gennai" to verify its usefulness for drafting administrative documents, conversational AI for staff and more (NTT DATA, March 23, 2026)
- Ibaraki City (Osaka) — Began a pilot of school-administration AI at municipal elementary and junior high schools with NTT West on October 13, 2026. An on-premises generative AI app supports searching and summarizing school documents and drafting texts (until March 31, 2027)
02Education and business
- Tokyo Online University — Decided to adopt tsuzumi 2 under the requirement of keeping student and staff data on campus, for class Q&A, support in creating teaching materials and exams, and course and career counseling (announced October 20, 2025)
- FUJIFILM Business Innovation — Began considering, with NTT DOCOMO BUSINESS, a generative AI solution that combines tsuzumi 2 with its technology for structuring documents (same announcement)
Published evaluations are mainly announcements by NTT and adopting organizations. On Japanese-language performance, NTT says tsuzumi 2 is among the world's best in its size class and holds its own against models several times larger (NTT's own evaluation on benchmarks it selected).
How it differs from other Japanese LLMs and LLM services
| Product | Provider | How you use it | Pricing (as stated on each product page at the time of checking) |
|---|---|---|---|
| tsuzumi 2 | NTT, Inc. | Deployed on-premises or in a private cloud as part of NTT Group companies' solutions | Not published (individual) |
| PLaMo翻訳 | Preferred Networks | A translation service based on the Japanese LLM "PLaMo", used on the web and in apps | Free plan available; paid plans are monthly |
| ELYZA | ELYZA (KDDI Group) | Generative AI adoption support and development of company- and industry-specific LLMs | Not published (individual quote) |
| さくらのAI Engine | SAKURA internet | An LLM inference API running in Japanese data centers | Free up to 3,000 requests per month, then pay-as-you-go |
| Ollama | Ollama Inc. (US) | A tool to run open models on your own PC | Local use is free |
tsuzumi 2 is an option for placing the model itself in your own environment and using it without sending data outside. If you want a Japanese LLM service that individuals can start using today, PLaMo翻訳 or さくらのAI Engine is easier. If you want outside help all the way to developing your own company-specific LLM, adoption support such as ELYZA is also worth comparing.
FAQ
Q. Can individuals use tsuzumi 2? A. There is no published service or free plan for individuals. NTT Group companies provide it individually to businesses and local governments.
Q. How much does tsuzumi 2 cost? A. Pricing is not published (as of October 2026). Because it depends on the configuration, you need to contact an NTT Group company.
Q. How is it different from the earlier "tsuzumi"? A. It increases the parameter count while keeping the lightness to run on one GPU, and strengthens Japanese performance, document search, summarization and information extraction, and knowledge of finance, healthcare and the public sector. The May 2026 update added the ability to read documents with charts and tables as images.
Q. Is it based on an overseas model? A. No. NTT says it is a fully domestic model built from scratch by NTT, without using other companies' openly available models as a base. NTT controls the training data itself and says it voluntarily removed data such as newspaper articles from the training data.
Q. Can it be used without sending confidential data to the cloud? A. Because it is a lightweight model that runs on one GPU, it can run on-premises or in a private cloud. NTT West offers an on-premises solution that runs it on a local network without going through the cloud.
The information on this page is based on what was published on NTT Group's official websites and press releases as of October 2026. Please check the official websites for the latest availability.
