이 모델의 강점은 무엇인가요?

Integrated with NTT Group's telecommunications infrastructure Potential for edge computing compatibility Optimized for Japanese language processing Continuous development backed by a major corporation's funding

이 모델의 약점은 무엇인가요?

Use limited to within NTT Group External availability undetermined Limited benchmark information Not open-source

어떤 용도에 가장 적합한가요?

AI solutions within NTT Group AI integrated with communication services Japanese language processing on edge devices Internal corporate chatbots

모델 목록으로

NTT독점

つづみ (tsuzumi)

Name: つづみ (tsuzumi)
Author: NTT

A domestic large-scale language model developed by the NTT Group. Specialized for Japanese, it is a concept to provide AI solutions deeply integrated with the NTT Group's communication infrastructure.

파라미터

7B〜13B

컨텍스트

라이선스

NTT License

출시일

2025-09-01

일본어 처리 능력

🇯🇵Native JP

Model developed by a Japanese company or specialized for Japanese. Highest Japanese understanding and generation capability.

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

・Integrated with NTT Group's telecommunications infrastructure
・Potential for edge computing compatibility
・Optimized for Japanese language processing
・Continuous development backed by a major corporation's funding

약점

・Use limited to within NTT Group
・External availability undetermined
・Limited benchmark information
・Not open-source

활용 사례

・AI solutions within NTT Group
・AI integrated with communication services
・Japanese language processing on edge devices
・Internal corporate chatbots

심층 분석

Parameter Size

30B (tsuzumi 2) / 7B (original tsuzumi)

Lightweight design for 1-GPU deployment

Japanese MT-Bench (writing)

8.6/10

Turn 1 evaluation; superior in Japanese humanities/writing

Hardware Requirement

1x NVIDIA A100 40GB

vs large models requiring multi-GPU clusters

Government Adoption

Selected for Digital Agency's Gennai platform

1 of 7 domestic LLMs for trial use

Market Demand

~2,000 inquiries in Japan

As of Q3 FY2025 (Oct-Dec 2025)

강점

・Exceptional Japanese language performance with deep understanding of business context and specialized terminology
・Ultra-lightweight design enabling on-premise/private cloud deployment with strong data sovereignty guarantees
・Efficient fine-tuning requiring 1/10th the training data of comparable models for domain specialization

약점

・Limited global benchmark visibility (no Arena Elo or SWE-Bench scores reported)
・No individual consumer offering; strictly B2B through NTT Group partnerships
・Multimodal capabilities (vision) still in development phase

경쟁사 비교

Model	Arena	SWE	GPQA	Price
tsuzumi 2 (NTT)	N/A	N/A	N/A	B2B licensing (est. hardware ~¥5M)
Gemma-3 27B (Google)	1470	68.2%	91.3%	$2.50/$7.50 per 1M tokens
Qwen-2.5 32B (Alibaba)	1465	72.1%	92.4%	$1.50/$6.00 per 1M tokens

개요

Tsuzumi is NTT's specialized Japanese large language model, designed primarily for enterprise and government applications requiring strict data sovereignty. The model has evolved from an initial 7-billion parameter version to the current 30-billion parameter 'tsuzumi 2' variant, maintaining its core philosophy of being lightweight enough to run on a single GPU while delivering top-tier Japanese language performance. Its development reflects Japan's strategic emphasis on domestic AI capabilities amid growing concerns about data security and reliance on foreign AI providers. The model's key innovation lies in its extreme efficiency: it achieves world-class Japanese comprehension with parameters an order of magnitude smaller than global frontier models. This enables on-premise deployment in sensitive sectors like finance, healthcare, and government, where data cannot leave controlled environments. NTT reports approximately 2,000 domestic inquiries as of late 2025, with the model being selected for Japan's Digital Agency 'Gennai' platform trial among seven domestic LLMs. While tsuzumi may not compete directly with trillion-parameter models in global benchmarks, it occupies a strategic niche as Japan's primary sovereign AI solution. Its performance on Japanese-specific benchmarks like MT-Bench shows particular strength in business writing and humanities, areas critical for domestic enterprise adoption. The recent addition of vision capabilities for processing charts and diagrams in business documents further enhances its practical utility in Japanese corporate environments.

벤치마크 및 성능

Tsuzumi's performance is primarily evaluated on Japanese-specific benchmarks rather than global competitions: | Benchmark/Task | tsuzumi 7B Score | Notes | |----------------|------------------|-------| | Japanese MT-Bench (writing) | 8.6/10 | Superior to many global models in Japanese contexts | | Japanese MT-Bench (humanities) | 8.45/10 | Strong cultural and contextual understanding | | Japanese MT-Bench (stem) | 7.6/10 | Technical understanding | | Japanese MT-Bench (coding) | 2.3/10 | Weakness in programming tasks | | Japanese MT-Bench (math) | 1.1/10 | Limited mathematical reasoning | The updated tsuzumi 2 (30B parameters) reportedly matches or exceeds GPT-5 in specific Japanese document understanding tasks, particularly: - Processing documents containing charts and graphs - Extracting structured data from Japanese business forms - Understanding business flows from diagrams - Handling numerical calculations in financial contexts However, the model lacks visibility in standard global benchmarks (Arena, SWE-Bench, GPQA) due to its specialized focus and limited international deployment.

상세 비교

Tsuzumi occupies a distinct niche compared to global LLMs: **vs Global Frontier Models (GPT-5, Claude, Gemini):** - **Advantages:** 10-20x lower hardware costs (single A100 vs multi-GPU clusters), guaranteed data sovereignty, superior Japanese business document understanding - **Disadvantages:** Limited multilingual capabilities, weaker global knowledge, lower performance on math/coding tasks - **Context Window:** 8k tokens (standard) vs 128k+ for frontier models **vs Other Japanese LLMs (ELYZA, Sarashina, cotomi):** - **Advantages:** NTT Group ecosystem integration, government platform adoption, established enterprise distribution - **Disadvantages:** More closed ecosystem; less community/open-source presence **Use Case Differentiation:** - Choose tsuzumi when: Handling confidential Japanese documents, requiring on-premise deployment, budget constraints exist for AI infrastructure - Choose global models when: Need multilingual support, complex coding assistance, or cutting-edge reasoning capabilities

커뮤니티 평가

Developer and researcher sentiment toward tsuzumi is mixed: **Enterprise/Government Sector:** - Strong interest from regulated industries (finance, healthcare, government) - Positive reception of data sovereignty guarantees - Adoption case studies in education (Tokyo University of Communication) and utilities (Chugoku Electric Power) **Technical Community:** - Skepticism among data scientists regarding performance parity with global models - Developer tests show tsuzumi 7B scoring 4% on SPI exam questions vs 77% for GPT-4o - Recognition of impressive efficiency given parameter constraints (7B performing comparably to much larger models on Japanese tasks) **Academic Perspective:** - NTT researchers emphasize the 'AI constellation' concept - specialized models working together rather than monolithic systems - Some view tsuzumi as primarily a strategic/marketing initiative rather than technological breakthrough **Adoption Pattern:** - Strong in traditional Japanese enterprises comfortable with NTT Group relationships - Limited international developer interest due to Japanese focus and B2B distribution model

활용 사례

1. **Confidential Document Processing in Financial Institutions:** Japanese banks and insurers can deploy tsuzumi on-premise to analyze contracts, extract data from forms (like tax documents), and automate compliance checks without exposing sensitive customer data to external clouds. The model's understanding of Japanese financial terminology and document layouts provides accuracy advantages over general-purpose models. 2. **Government Document Automation:** As seen in the Digital Agency's Gennai platform adoption, tsuzumi can assist civil servants in drafting and summarizing administrative documents, understanding complex regulations, and processing citizen inquiries while maintaining strict data sovereignty requirements mandated for government operations. 3. **Manufacturing Knowledge Management:** Companies like Kawasaki Heavy Industries use tsuzumi to extract and systematize tacit knowledge from veteran engineers. The model can process Japanese technical documentation, maintenance manuals, and safety protocols to create searchable knowledge bases for workforce training. 4. **Healthcare Record Analysis (Future Potential):** With ongoing vision capability development, tsuzumi could process handwritten medical charts, prescription forms, and diagnostic images while keeping patient data within hospital networks, addressing privacy regulations in Japan's healthcare system. **When to choose over alternatives:** Select tsuzumi when data must remain within organizational boundaries, when primary users are Japanese speakers, when hardware budgets are constrained, and when the use case involves typical Japanese business documents rather than cutting-edge AI research or multilingual applications.