Tsuzumi
Lightweight Japanese LLM for business use
NTT developed tsuzumi, a family of lightweight large language models focused on Japanese processing and corporate efficiency. Released in 2024 with an updated version in 2025, it runs on modest hardware to lower costs and support on-premises data security.
- Runs on a single GPU or CPU
- Developed entirely from scratch by NTT
- Optimized for Japanese business tasks
- Supports on-premises deployment
- Released commercially in March 2024
What it is
tsuzumi is a series of large language models built by NTT for Japanese language tasks. The initial release included two sizes: an ultra-lightweight version with 600 million parameters and a standard version with 7 billion. These models are designed to run efficiently on single GPUs or even CPUs, reducing the energy and hardware costs typical of larger systems. In October 2025, NTT launched tsuzumi 2, which improves performance on business-specific tasks while maintaining low operational requirements. The model targets corporate users who need secure, on-premises solutions for document handling and customer service.
How it came about
NTT announced tsuzumi in November 2023, with commercial services starting in March 2024. The development drew on over forty years of natural language processing research at NTT Laboratories. Early trials involved Kyoto University Hospital for medical record structuring and Tokio Marine & Nichido Fire Insurance for contact centre automation. The company positioned the model as a sustainable alternative to energy-intensive large-scale systems. In 2025, NTT released tsuzumi 2, claiming it matches the performance of much larger models on specific benchmarks while remaining lightweight.
Inside the model
The model uses a decoder-only transformer architecture. NTT developed it from scratch without using existing open-source foundations, aiming to control training data quality and comply with Japanese laws. This approach addresses legal concerns about copyright and data usage. The system supports adapter-based tuning, allowing users to add industry-specific knowledge without retraining the entire model. Tsuzumi 2 focuses on instruction following, ensuring outputs meet strict formatting needs like JSON extraction. While current versions handle text, NTT has outlined plans for multimodal features involving vision and audio.
What it can do
Tsuzumi performs well on Japanese language benchmarks, including MT-bench, where NTT claims it rivals larger models like GPT-5 in certain tasks. It excels in question answering, document summarization, and information extraction. The model is optimized for business scenarios, such as generating reports in specific styles or parsing source code. Tsuzumi 2 shows significant improvements over its predecessor in these practical applications. Its lightweight design allows it to operate on older hardware, making it accessible for companies with limited computing resources.
How to use it
NTT offers tsuzumi through its group companies as a commercial service for corporate clients. It is available via API and can be integrated into custom systems. The model supports on-premises deployment, which helps organisations keep sensitive data within their own infrastructure. This setup avoids the recurring costs of public API usage and reduces liability risks associated with external data processing. Licences are proprietary, reflecting the company’s full-scratch development strategy.
Safety and testing
NTT emphasizes compliance with domestic Japanese laws and regulations. By building the model from scratch, the company controls the training datasets to minimize biases and ensure rights clearance. This method aims to prevent legal issues related to unauthorized data use, such as those seen in other industry lawsuits. The focus on on-premises deployment also enhances security for corporate users handling private information.
Reception and debate
The model was presented as a solution for sustainability, addressing the high power consumption of typical large language models. NTT highlighted its potential to reduce operational costs and improve employee experience in sectors like medicine and insurance. The release of tsuzumi 2 reinforced its position as a high-performance option for businesses seeking efficient, locally developed AI tools.
Why it matters
Tsuzumi demonstrates that smaller, specialized models can compete with larger general-purpose systems in specific languages and tasks. Its success in medical and insurance trials suggests a viable path for industry-specific AI adoption. By prioritizing efficiency and legal compliance, NTT has created a model that appeals to enterprises concerned with cost and data sovereignty.
Questions people ask about Tsuzumi
What makes tsuzumi different from other large language models?
Tsuzumi is designed to be lightweight and efficient, running on modest hardware like a single GPU or CPU. It focuses specifically on Japanese language processing and corporate use cases, offering lower costs and better data security through on-premises deployment.
Can tsuzumi handle multiple languages?
The model is primarily optimized for Japanese. While it may have some multilingual capacity, its main strength and design focus lie in processing Japanese text for business applications such as summarization and information extraction.
Is tsuzumi open source?
No, tsuzumi is a proprietary model developed by NTT. It is offered as a commercial service to corporate customers, often with options for on-premises integration, but the weights are not publicly available for download.
Good reading and sources
There is no English Wikipedia article on Tsuzumi yet, so the facts on this page come from the sources listed above: rd.ntt/e/research/LLM_tsuzumi.html, group.ntt/en/magazine/blog/tsuzumi/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.