NTT tsuzumi
Japan's sovereign lightweight LLM — enterprise-grade Japanese AI on a single H100, built from scratch by NTT
Overview
NTT tsuzumi is a family of lightweight large language models developed entirely in-house by Nippon Telegraph and Telephone (NTT), drawing on the company's 40-plus years of natural language processing research. The original model launched commercially in March 2024 with 600 million and 7 billion parameter variants designed to run on minimal hardware — even a single CPU or one GPU — making high-quality Japanese-language AI accessible without the massive infrastructure costs of frontier models. tsuzumi 2, released in October 2025 and updated in May 2026, extended the family to approximately 30 billion parameters while retaining single-GPU (H100) inference, adding multimodal support for charts, tables and document images, and deepening domain specialisation in finance, healthcare, and public administration.
What distinguishes tsuzumi from global competitors is its "purely domestic" provenance: trained from scratch on data NTT owns or licenses under Japanese law, with no reliance on overseas foundation models. This positions it as a sovereign AI option for organisations with strict data-residency, copyright-compliance, or national-security requirements. It is deployed on-premises, in private clouds, or via Microsoft Azure's Model-as-a-Service (MaaS) serverless API. Verified enterprise adopters include Tokyo Online University (campus-network deployment), NTT DOCOMO BUSINESS and FUJIFILM Business Innovation (knowledge management), Mie University Hospital (clinical note summarisation), and Japan's national Digital Agency, which selected tsuzumi 2 for its "GENAI" government-wide AI trial involving approximately 180,000 civil servants from fiscal year 2026.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- No public rate card
- Sign-up required for pricing
- On-prem, private cloud, or Azure MaaS deployment
Use cases
What you can produce with NTT tsuzumi
- Domain-adapted Japanese-language chatbot fine-tuned on your own internal documents using adapter tuning with minimal labelled data
- Automated extraction of structured data (figures, tables, credit-screening fields) from scanned or image-embedded Japanese business documents
- RAG-powered internal knowledge-base Q&A system deployed entirely within your corporate network or private cloud
- Summarised nursing notes or clinical documentation generated on-premises with no patient data leaving the hospital environment
- Serverless API integration via Microsoft Azure AI Foundry for token-billed Japanese-English bilingual text generation in enterprise applications
- Fine-tuned public-sector chatbot handling citizen inquiries against local government ordinances and procedural documentation
- Multimodal document intelligence pipeline that ingests PDF reports with charts and outputs structured insights or draft summaries
ASEAN Perspective
NTT tsuzumi in Southeast Asia
NTT tsuzumi's ASEAN relevance is primarily indirect but growing. As a "purely domestic Japanese LLM" built under Japanese law with on-premises deployment support, it appeals to Southeast Asian enterprises with Japan operations or supply-chain ties that share Japanese data-sovereignty concerns — a live issue for financial institutions, government agencies, and healthcare providers across the region. NTT's parallel investment in a 8,100 km Japan-ASEAN subsea cable (connecting Japan, Malaysia, Singapore, South Korea, the Philippines, and Taiwan) signals that the group intends to extend its AI infrastructure footprint into the region, which could eventually bring tsuzumi closer to ASEAN data-residency compliance needs. At present, multimodal and multilingual support for ASEAN languages (Bahasa Melayu, Bahasa Indonesia, Thai, Vietnamese) is not in the public roadmap, making tsuzumi 2 a watch-and-wait option for most Southeast Asian organisations rather than a production-ready ASEAN AI platform.
NTT tsuzumi 2 is a credible sovereign-AI play, not just a brand exercise. The single-H100 inference claim checks out across multiple independent reports, and being tapped for Japan's Digital Agency "Gennai" trial — rolling out to roughly 180,000 civil servants from fiscal 2026 — is real institutional validation, not just NTT's own press release talking. For Japanese-language enterprise work — document Q&A, RAG over internal manuals, extracting numbers from financial tables and charts — NTT's own benchmarks show it beating comparably-sized models on Rakuda and posting an 81.3% win rate against GPT-3.5.
The catch is access. There's no public rate card — Azure Marketplace pricing only appears after signup — no open-source release, and no self-serve trial, so evaluating it means going through NTT or an Azure MaaS sales conversation first. Benchmarks are also NTT's own; independent replication is thin. Outside Japan, this isn't the model to reach for.
What people say
There's no G2 page, no Capterra listing, no app-store rating for NTT tsuzumi 2 — it's sold through NTT relationships and Azure's Model-as-a-Service marketplace, not to individual developers. That alone tells you who this is built for: large Japanese enterprises and government bodies, not startups shopping for an API key.
The single-GPU story is the real differentiator. tsuzumi 2 runs inference on one H100, which NTT and independent reports both confirm, and that translates into materially lower hosting costs than comparable frontier models for Japanese-heavy workloads — document Q&A, RAG over internal manuals, extracting figures from financial documents with charts and tables. Japan's Digital Agency selected it for its "Gennai" government AI trial covering roughly 180,000 civil servants starting fiscal 2026, which is a meaningful vote of institutional confidence beyond NTT's own marketing.
Two things temper the enthusiasm. First, NTT's benchmark claims — an 81.3% win rate against GPT-3.5, strong Rakuda scores — come from NTT itself; there's no independent lab replicating these numbers yet. Second, pricing is genuinely opaque: there's no public rate card, and even Azure Marketplace token pricing only shows up after you sign up for access. No open-source release, no public sandbox, no free tier to poke at before committing.
For a Japanese enterprise with data-residency requirements and existing NTT or Azure relationships, tsuzumi 2 is a legitimate, low-cost option worth evaluating. For anyone outside that lane — non-Japanese-language workloads, ASEAN teams without an NTT relationship, developers wanting to test-drive an API — this isn't built for you, and NTT doesn't seem to be chasing that market anyway.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- The name 'tsuzumi' (鼓) is taken from a traditional Japanese hourglass-shaped hand drum — chosen to reflect the model's compact yet resonant design philosophy.
- The 600 million parameter variant of the original tsuzumi can run inference entirely on a CPU, making it roughly 300 times lighter than GPT-3's 175 billion parameters.
- Japan's national Digital Agency selected tsuzumi 2 for a GENAI government AI pilot reaching approximately 180,000 civil servants across every ministry from fiscal year 2026.
- NTT trained tsuzumi from scratch using only data it owns or has licensed — a deliberate choice to avoid the copyright-scraping controversies that have dogged other LLM developers.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including NTT tsuzumi's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with NTT tsuzumi unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to NTT tsuzumi directly →
Spotted something out of date? Suggest an update →
NTT tsuzumi in the news
Alternatives to NTT tsuzumi
More in LLMs & Chat
