By NTTLearning Objectives
- Understand what tsuzumi 2 is and why Japan is investing in domestic language models
- Evaluate the model's single-GPU efficiency and government deployment
- Assess tsuzumi 2's role in Japan's sovereign AI strategy
What Is tsuzumi 2?
tsuzumi 2 is a 30 billion parameter language model from NTT (Nippon Telegraph and Telephone), Japan's largest telecom company. Released in October 2025, it is optimized specifically for Japanese language tasks and designed for enterprise and government deployment — not consumer chatbots.
The model's standout feature: it runs on a single NVIDIA H100 GPU (~$35,000 in hardware). This makes it dramatically more affordable and energy-efficient to deploy than frontier models that require clusters of 8 or more GPUs, while still achieving what NTT claims are world-top results for Japanese among models of comparable size.
In March 2026, tsuzumi 2 was selected for Japan's Gennai generative AI platform — the government's initiative to bring AI capabilities to approximately 180,000 government staff, with broad rollout expected from May 2026.
💡Key Concept
Single-GPU Deployment: Most frontier AI models require clusters of 4 to 8 high-end GPUs costing $100,000 or more. A model that runs on a single GPU dramatically reduces deployment cost, power consumption, and operational complexity — making AI practical for organizations that cannot justify or afford large GPU clusters. This is especially important for on-premises deployment where data sovereignty requires keeping AI systems within organizational control.
Key Specifications
| Spec | Detail |
|---|---|
| Parameters | 30 billion |
| Hardware Requirement | Single NVIDIA H100 GPU |
| Language Optimization | Japanese (world-top results at comparable model size) |
| Specialized Training | Finance, municipal government, and healthcare domains |
| Government Selection | Japan Gennai platform (180,000 government staff) |
| License | Commercial/enterprise (not open-source) |
| Previous Version | tsuzumi 1 (7 billion parameters) |
Government Deployment: Gennai
Japan's Digital Agency selected tsuzumi 2 as one of 7 AI models for the Gennai generative AI platform — the government's plan to bring AI capabilities to its workforce:
| Vendor | Model | Notes |
|---|---|---|
| NTT | tsuzumi 2 | 30 billion parameter Japanese-optimized model |
| KDDI/ELYZA | Llama-3.1-ELYZA-JP-70B | Japanese fine-tuned Llama model |
| PFN | PLaMo 2.0 Prime | Preferred Networks' flagship model |
| NEC | cotomi v3 | NEC's enterprise Japanese model |
Broad rollout is expected from May 2026, serving approximately 180,000 government employees across Japan's national agencies.
Domain Specialization
Unlike general-purpose models, tsuzumi 2 was extensively trained on domain-specific data for three key sectors:
- Finance — financial document analysis, regulatory compliance, reporting
- Municipal government — citizen services, document processing, administrative workflows
- Healthcare — medical documentation, clinical decision support, patient communication
This specialized training makes tsuzumi 2 more accurate for these domains than general-purpose models that treat all topics equally.
tsuzumi 2 vs. Japanese AI Competitors
| Model | Company | Parameters | Key Difference |
|---|---|---|---|
| tsuzumi 2 | NTT | 30 billion | Single-GPU deployment; government-selected; domain-specialized |
| PLaMo 2.0 Prime | PFN (Preferred Networks) | Not disclosed | Strong research heritage; Gennai-selected |
| cotomi v3 | NEC | Not disclosed | NEC's enterprise focus; Gennai-selected |
| ELYZA-JP-70B | KDDI/ELYZA | 70 billion | Llama-based; stronger general performance; larger model |
Company Details
| Detail | Info |
|---|---|
| Company | NTT Inc. (formerly Nippon Telegraph and Telephone; renamed July 2025) |
| Founded | 1952 |
| Headquarters | Tokyo, Japan |
| Revenue | ~13.70 trillion yen (~$91 billion; fiscal year ending March 2025) |
| Employees | ~340,000 |
| Global Ranking | Fortune Global 500 number 128; 6th largest telecom by revenue |
| Website | ntt.com |
Strengths
- Single-GPU efficiency — runs on one H100 ($35,000), making it dramatically cheaper and simpler to deploy than multi-GPU models
- Japanese language excellence — world-top results among comparable-size models for Japanese tasks
- Government validation — selected for Japan's Gennai platform serving 180,000 staff
- Domain specialization — trained specifically for finance, government, and healthcare use cases
- NTT backing — $91 billion revenue telecom giant provides enterprise trust and long-term support
Limitations and Considerations
- Not open-source — commercial/enterprise product; cannot be freely downloaded or modified like Llama or Mistral
- Japanese-focused — optimized specifically for Japanese; not competitive for English or multilingual tasks
- Enterprise-only — no consumer-facing chatbot; designed for B2B and B2G deployment
- Benchmark transparency — "world-top results for Japanese" is NTT's own claim; Japanese language AI benchmarks are less standardized than English ones
- Niche market — targets a specific geographic and linguistic market; limited applicability outside Japan
Key Takeaways
- tsuzumi 2 is NTT's 30 billion parameter Japanese-optimized model — running on a single H100 GPU for cost-efficient enterprise and government deployment
- Selected for Japan's Gennai government AI platform (180,000 staff, rollout from May 2026) alongside models from KDDI, PFN, and NEC
- Specialized training in finance, municipal government, and healthcare makes it more accurate for these domains than general-purpose models
- Backed by NTT ($91 billion revenue, 340,000 employees) — enterprise-grade support but not open-source
