# Tsuzumi (model family)
Tsuzumi is a family of large language models for Japanese and English developed by NTT, Inc., the Japanese telecommunications and technology group, and announced on November 1, 2023.[1](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html) The family is built around lightweight models: a 0.6-billion-parameter version that runs on a CPU and a 7-billion-parameter version that performs high-speed inference on a single GPU, sizes NTT presents as 1/300 and 1/25 of GPT-3's 175 billion parameters respectively.[2](https://group.ntt/en/magazine/blog/tsuzumi/) A second generation, tsuzumi 2, was released on October 20, 2025 and updated on May 19, 2026.[3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) • [4](https://group.ntt/en/newsrelease/2026/05/19/260519a.html) This article covers the model family only; NTT as a company and any products built on the models are separate subjects.
| Fact | Detail |
|---|---|
| Maker | NTT, Inc. (vendor-reported) |
| First announcement | November 1, 2023; pre-training began around June 2023[1](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html) |
| First-generation versions | 0.6B (CPU) and 7B (single GPU), commercial launch March 25, 2024[5](https://www.rd.ntt/e/research/LLM_tsuzumi/ntt_llm_press_conference_slides_EN.pdf) • [1](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html) |
| Second generation | tsuzumi 2, October 20, 2025; update May 19, 2026[3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) • [4](https://group.ntt/en/newsrelease/2026/05/19/260519a.html) |
| Languages | Japanese and English; the 2023 roadmap also listed 21 other languages and programming languages for the 7B model[5](https://www.rd.ntt/e/research/LLM_tsuzumi/ntt_llm_press_conference_slides_EN.pdf) |
| Hardware footprint | CPU for 0.6B; single GPU for 7B and tsuzumi 2, for on-premises or private-cloud operation[2](https://group.ntt/en/magazine/blog/tsuzumi/) • [3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) |
| Availability | Generally available in the Microsoft Azure AI Foundry model catalog as tsuzumi-7b[6](https://ai.azure.com/catalog/models/tsuzumi-7b/) |
## Release timeline and versions
Pre-training started around June 2023, and NTT announced tsuzumi on November 1, 2023.[1](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html) The 2023 roadmap set out an ultra-lightweight 0.6B model for CPU (Japanese only, planned March 2024), a lightweight 7B model for low-grade GPU (trial from October 2023, release March 2024, covering Japanese, English, 21 other languages, and programming languages), and larger 13B-plus models after April 2024.[5](https://www.rd.ntt/e/research/LLM_tsuzumi/ntt_llm_press_conference_slides_EN.pdf) Commercial service was announced on March 25, 2024.[1](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html)
The next documented release is tsuzumi 2 on October 20, 2025, which NTT describes as a lightweight Japanese-processing LLM aimed at reducing the electricity consumption and cost of running large models.[3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) On May 19, 2026 NTT announced an update adding visual understanding of charts and tables in Japanese business documents, processed as images, within a single-GPU environment, along with stronger logical reasoning including numerical processing such as calculations on sales figures and interpretation of API documentation.[4](https://group.ntt/en/newsrelease/2026/05/19/260519a.html) NTT says the updated model will be rolled out through companies across the NTT Group.[4](https://group.ntt/en/newsrelease/2026/05/19/260519a.html)
## Architecture and training as published
NTT has published the following about the first generation. The 0.6B model has 600 million parameters and runs on a CPU; the 7B model has 7 billion parameters and performs high-speed inference on one GPU.[2](https://group.ntt/en/magazine/blog/tsuzumi/) Pre-training used more than 1,000 billion tokens spanning domains from technical documents to entertainment, drawn from in-house data mainly in Japanese and English plus code and multilingual data.[1](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html)
Two Japanese-specific choices are central to NTT's account. First, NTT built a custom tokenizer whose vocabulary reflects [Japanese grammar](https://www.edgepedia.org/japanese-grammar), lexical properties, and frequently occurring compound words, which it says improves token efficiency in Japanese.[7](https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr202608fa2.html) Second, it developed a filtering algorithm to extract high-quality Japanese text from a massive web corpus, removing noise to raise data purity, which it credits for accurate knowledge about Japan.[7](https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr202608fa2.html) Alignment work focused on mathematical tasks, a domain NTT treats as representative for building reasoning ability in a compact model.[7](https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr202608fa2.html) For tsuzumi 2, NTT states the model was built strictly in-house from the ground up, not adapted from overseas models, using closely controlled training data.[8](https://www.global.ntt/insights-hub/ntts-tsuzumi-2/)
Several technical details are not disclosed in the record: the context length, the training infrastructure, the detailed data mix, and the parameter count and architecture of tsuzumi 2, which NTT describes only as lightweight and single-GPU.[3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) A third-party directory claims tsuzumi 2 extended the family to approximately 30 billion parameters with single-GPU (H100) inference, but NTT's own releases state no parameter count, so this figure is unverified and the two accounts are unresolved.[9](https://recatools.com/ai-directory/ntt-tsuzumi/)
## Benchmark performance: vendor claims versus independent evidence
Every performance figure in the public record for Tsuzumi is vendor-reported. At the 2023 launch NTT claimed the 7B model achieved top-level accuracy in Japanese task evaluation, built on more than 40 years of NTT natural language processing research.[10](https://www.rd.ntt/forum/2023/doc/P01_leaf_e.pdf) NTT reported that tsuzumi-7B surpassed GPT-3.5 on the Rakuda benchmark, which uses 40 questions on Japanese geography, politics, history, and society scored by GPT-4, and that it significantly outperformed domestic LLMs of the same class.[5](https://www.rd.ntt/e/research/LLM_tsuzumi/ntt_llm_press_conference_slides_EN.pdf) • [1](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html) On English, NTT claimed roughly the same performance as [Llama 2](https://www.edgepedia.org/llama-2) 7B on a 10-task lm-evaluation-harness common-sense average.[5](https://www.rd.ntt/e/research/LLM_tsuzumi/ntt_llm_press_conference_slides_EN.pdf) For tsuzumi 2, NTT claims world-top Japanese results among models of comparable size and business-domain performance on par with or exceeding models several times larger, in knowledge, analysis, instruction-following, and safety.[3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) • [7](https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr202608fa2.html)
The Azure catalog's vendor-reported Japanese MT-bench turn-1 scores for tsuzumi-7B show the model's profile: writing 8.6, STEM 7.6, humanities 8.45, roleplay 6.3, extraction 5.6, but coding 2.3, math 1.1, and reasoning 2.1 in Japanese, with generally lower English scores (extraction 2.9, reasoning 4.3).[6](https://ai.azure.com/catalog/models/tsuzumi-7b/) These numbers indicate markedly weaker math, coding, and reasoning than writing and humanities, consistent with a small model despite NTT's math-focused alignment.
No independent evaluation appears in the record: the sources contain no third-party JGLUE, Japanese MT-Bench, or LLM-jp results, and no comparison data against Qwen, ELYZA, [Swallow](https://www.edgepedia.org/swallow-ai-model), PLaMo, or Sarashina. Readers should treat all performance claims, including the "world-top" and GPT-3.5 comparisons, as NTT's own until independent measurements exist.
## Adoption and availability
tsuzumi-7b is generally available in the Microsoft Azure AI Foundry model catalog, described as a lightweight NTT LLM for Japanese and English instruction following using a transformer architecture with supervised fine-tuning.[6](https://ai.azure.com/catalog/models/tsuzumi-7b/) Before the March 2024 commercial launch, trials ran at Kyoto University Hospital in the medical field and at Tokio Marine & Nichido Fire Insurance in the contact center.[2](https://group.ntt/en/magazine/blog/tsuzumi/) In October 2025, Tokyo Online University adopted tsuzumi 2 for an on-premise platform that keeps student and staff data within the [campus network](https://www.edgepedia.org/campus-network), and NTT DOCOMO BUSINESS partnered with FUJIFILM Business Innovation to explore a generative-AI solution built on tsuzumi 2.[3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) NTT says the lightweight design has led to wide adoption for sensitive-information use cases in on-premises and private cloud environments.[4](https://group.ntt/en/newsrelease/2026/05/19/260519a.html)
The sources do not state Tsuzumi's license terms, who may use it commercially, or how its licensing compares with Llama or [Qwen](https://www.edgepedia.org/qwen-model-family), and they do not document whether the weights were openly released, for example on [Hugging Face](https://www.edgepedia.org/hugging-face). The Azure catalog listing is the only confirmed third-party availability channel in the record.[6](https://ai.azure.com/catalog/models/tsuzumi-7b/)
## Tsuzumi and Japan's sovereign-AI strategy
NTT frames tsuzumi as a sovereign LLM, citing the Japanese government's AI Basic Plan, which works to strengthen the development of sovereign AI (AI developed and controlled within a nation's borders).[7](https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr202608fa2.html) The commercial logic is closed-data demand: NTT reports that more than 60% of NTT Group companies' customers request that personal information and highly confidential data be used in a closed environment, which a lightweight model that runs on a single GPU on-premises or in a private cloud is designed to serve.[7](https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr202608fa2.html) • [3](https://group.ntt/en/newsrelease/2025/10/20/251020a.html) This explains why a telecom company built its own model family: its customer base is the market for domestically developed, closed-environment AI. The sources do not cover any Tsuzumi involvement in GENIAC, ABCI, or Fugaku.
## References
1. [NTT's Large Language Model "tsuzumi" (NTT Technical Review, Aug 2024)](https://ntt-review.jp/archive/ntttechnical.php?contents=ntr202408fr1_s.html)
2. [NTT's Large Language Model "tsuzumi" is Here! (NTT STORY)](https://group.ntt/en/magazine/blog/tsuzumi/)
3. [NTT's Next-Generation LLM "tsuzumi 2" Now Available](https://group.ntt/en/newsrelease/2025/10/20/251020a.html)
4. [NTT's LLM tsuzumi 2 Updated](https://group.ntt/en/newsrelease/2026/05/19/260519a.html)
5. [Overview of NTT's LLM (press conference slides)](https://www.rd.ntt/e/research/LLM_tsuzumi/ntt_llm_press_conference_slides_EN.pdf)
6. [Microsoft Foundry Model Catalog: tsuzumi-7b](https://ai.azure.com/catalog/models/tsuzumi-7b/)
7. [tsuzumi: Challenges in Developing a Sovereign Large Language Model (NTT Technical Review, 2026)](https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr202608fa2.html)
8. [tsuzumi 2: Secure, Efficient AI for Enterprise](https://www.global.ntt/insights-hub/ntts-tsuzumi-2/)
9. [NTT tsuzumi Review (recatools directory)](https://recatools.com/ai-directory/ntt-tsuzumi/)
10. [NTT R&D Forum 2023: Large language model "tsuzumi" (leaflet)](https://www.rd.ntt/forum/2023/doc/P01_leaf_e.pdf)
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Model families and named models › Large language model families › Open-weight model families*
*Initially written Sep 17, 2026 · Reviewed: Sep 19, 2026 · Edited: Sep 19, 2026; Oct 11, 2026; Oct 25, 2026 · Last review: Sep 19, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license