As ChatGPT captured global attention, a quieter revolution was unfolding in Japanese research labs. Fujitsu, NEC, NTT, and a wave of startups have been building large language models designed not just to understand Japanese — but to reason about Japanese business, culture, and regulatory frameworks. In a world where AI sovereignty is becoming a strategic imperative, Japan’s homegrown LLMs may matter far more than their international visibility suggests.
Why Japan Needs Sovereign AI
The argument for domestically developed large language models extends well beyond national pride. Japanese — with its three writing systems (hiragana, katakana, and kanji), complex honorific structures, and context-dependent grammar — has historically been underrepresented in the training data of English-centric foundation models. When GPT-4 launched, estimates suggested that Japanese text constituted less than 2% of its training corpus despite Japan being the world’s third-largest economy.
This underrepresentation creates practical problems. English-centric models often struggle with keigo (formal Japanese), industry-specific terminology, and the implicit contextual understanding that Japanese business communication requires. They may mishandle the nuances of Japanese legal documents, misinterpret the polite indirectness of business correspondence, or produce outputs that are grammatically correct but culturally inappropriate. For Japanese enterprises deploying AI in customer-facing or decision-critical applications, these limitations represent unacceptable risks.
There are also concerns about data sovereignty. Japanese corporations, government agencies, and healthcare institutions are reluctant to send sensitive data to foreign AI providers — a concern amplified by regulatory requirements under Japan’s Act on the Protection of Personal Information (APPI) and the Economic Security Promotion Act. Domestically developed and hosted LLMs offer a path to AI adoption that satisfies these regulatory and sovereignty requirements.
The Linguistic Challenge
Building effective LLMs for Japanese presents unique technical challenges that extend beyond simply adding Japanese data to training sets. The tokenization problem is fundamental: models designed for English typically use byte-pair encoding (BPE) tokenizers that fragment Japanese text inefficiently, requiring three to five times more tokens to represent the same semantic content as English. This inefficiency increases computational costs and can degrade model performance on Japanese tasks.
Japanese also presents challenges in knowledge representation. Much of the world’s AI training data reflects English-language knowledge structures — Western legal frameworks, American business practices, European cultural references. A model that genuinely serves Japanese users must be trained on data that captures Japanese institutional knowledge, regulatory frameworks, and cultural context. This requires curating large, high-quality Japanese corpora — a resource-intensive undertaking that several Japanese organizations have now completed.
The Major Players
Fujitsu: Takane and Kozuchi
Fujitsu has approached the LLM challenge through its Kozuchi AI platform, which integrates multiple AI capabilities including its Takane large language model. Takane, developed in collaboration with international partners and trained on carefully curated Japanese datasets, is designed specifically for enterprise deployment. Fujitsu has positioned Takane not as a general-purpose chatbot competitor but as a domain-adaptable foundation model that enterprises can fine-tune for specific use cases — legal document analysis, customer service, technical documentation, and regulatory compliance.
Fujitsu’s approach reflects a strategic calculation that the enterprise AI market will favor specialized, controllable models over general-purpose giants. Takane is offered through Fujitsu’s consulting and system integration channels, embedded within industry-specific solutions that include data preparation, fine-tuning, safety evaluation, and ongoing model management. This full-stack approach leverages Fujitsu’s existing relationships with Japan’s largest enterprises and government agencies.
NEC: cotomi
NEC developed cotomi, a family of LLMs that has achieved notable benchmarks in Japanese language understanding. The cotomi models, ranging from lighter variants suitable for edge deployment to larger models for complex reasoning tasks, were trained on a proprietary Japanese corpus that NEC claims is among the largest ever assembled for Japanese-language AI training.
NEC has emphasized cotomi’s performance on Japanese-specific evaluation benchmarks, where it has demonstrated competitive or superior results compared to much larger international models on tasks involving Japanese reading comprehension, legal text analysis, and business document summarization. The company has integrated cotomi into its NEC Digital Platform, making it available to enterprise customers through both API access and on-premises deployment — a critical capability for customers with strict data residency requirements.
NTT: tsuzumi
NTT unveiled tsuzumi in late 2023, a lightweight LLM explicitly designed for efficiency and Japanese language mastery. The model’s name — derived from a traditional Japanese hand drum — reflects NTT’s positioning of tsuzumi as a distinctly Japanese AI. What distinguishes tsuzumi is its parameter efficiency: the model achieves competitive Japanese language performance with significantly fewer parameters than comparable international models, reducing both the computational cost and the carbon footprint of deployment.
NTT’s strategy with tsuzumi centers on telecommunications and enterprise applications where the company’s existing infrastructure provides deployment advantages. The model is integrated with NTT’s cloud and network services, enabling use cases such as automated customer support, network operations intelligence, and document analysis. NTT has also emphasized tsuzumi’s multimodal capabilities, including Japanese speech recognition and generation — leveraging the company’s extensive research in speech technology.
Preferred Networks: PLaMo
Preferred Networks (PFN), Japan’s most prominent deep learning startup, entered the LLM arena with PLaMo, a family of open-weight models that has attracted significant attention from the research community. PFN, which built its reputation on deep learning applications in manufacturing and robotics, brings substantial computational infrastructure and research talent to the LLM challenge.
PLaMo’s open-weight approach distinguishes it from the proprietary models offered by Fujitsu, NEC, and NTT. By making model weights available to researchers and developers, PFN has catalyzed a community of Japanese-language AI development that extends beyond what any single organization could achieve. The company has released multiple model sizes and fine-tuned variants, including instruction-following models and models optimized for specific domains.
Sakana AI: The Research Frontier
Sakana AI, founded in 2023 by former Google Brain researchers David Ha and Llion Jones (a co-author of the landmark “Attention Is All You Need” transformer paper), has quickly become one of the most closely watched AI startups globally. Based in Tokyo, the company has raised substantial funding and is pursuing a distinctive research agenda focused on nature-inspired AI architectures — evolutionary algorithms, swarm intelligence, and model merging techniques that combine multiple specialized models into more capable systems.
Sakana’s presence in Tokyo — rather than Silicon Valley — represents a deliberate bet on Japan’s AI ecosystem. The company has cited Japan’s technical talent, research culture, and proximity to the large enterprise customers who will ultimately deploy AI at scale. Sakana’s model merging research has produced results that challenge assumptions about how LLMs need to be trained, suggesting that smaller, specialized models can be combined to achieve performance comparable to much larger monolithic models.
ABEJA: Enterprise AI Pioneer
ABEJA, one of Japan’s earliest AI startups, has developed its own LLM capabilities and integrated them into its enterprise AI platform. The company’s approach emphasizes practical deployment over benchmark performance, focusing on the end-to-end challenges of making LLMs useful in Japanese enterprise environments — data preparation, model customization, safety evaluation, and integration with existing business systems.
The Competitive Landscape
| Model/Company | Parameters | Architecture | Deployment | Key Differentiator |
|---|---|---|---|---|
| Takane (Fujitsu) | Multiple sizes | Transformer | Cloud / On-prem | Enterprise integration via Kozuchi |
| cotomi (NEC) | Multiple sizes | Transformer | Cloud / On-prem | Japanese benchmark leadership |
| tsuzumi (NTT) | 7B / 70B | Transformer | NTT Cloud | Parameter efficiency, multimodal |
| PLaMo (PFN) | 13B / 100B | Transformer | Open weights | Open-source community ecosystem |
| Sakana AI | Various merged | Evolutionary | API / Research | Model merging, nature-inspired AI |
| ABEJA LLM | Multiple sizes | Transformer | Cloud / On-prem | Enterprise deployment expertise |
Sources: Company announcements and technical reports (2023-2025); Nikkei BP AI market analysis; Stanford HAI AI Index Report (2024)
Enterprise Adoption: Where Japanese LLMs Are Being Deployed
The deployment of Japanese LLMs is accelerating across several sectors where the advantages of domestic models are most pronounced.
In financial services, Japanese banks and insurance companies are deploying domestic LLMs for regulatory document analysis, compliance checking, and customer communication. The financial sector’s stringent data residency requirements and the complexity of Japanese financial regulations make domestic models particularly attractive. Mizuho Financial Group, MUFG, and Sumitomo Mitsui Financial Group have all announced LLM initiatives utilizing Japanese-developed models.
In government and public services, the Digital Agency has been evaluating domestic LLMs for public-sector applications including citizen inquiry handling, policy document drafting, and administrative process automation. The government’s emphasis on digital sovereignty — articulated in the Economic Security Promotion Act — creates a strong preference for domestically developed AI technologies in sensitive applications.
In manufacturing, Japanese LLMs are being deployed for technical documentation management, quality control reporting, and supply chain communication. Toyota, Honda, and other major manufacturers have explored domestic LLM solutions for internal knowledge management — applications where the models must understand not only Japanese but the specific technical vocabularies and communication patterns of Japanese manufacturing.
In healthcare, the complexity of Japanese medical terminology and the sensitivity of patient data have driven interest in domestic LLMs for clinical documentation, medical literature analysis, and patient communication. Several university hospitals have initiated pilot programs evaluating Japanese LLMs for these applications.
METI AI Guidelines and the Regulatory Framework
The Ministry of Economy, Trade and Industry (METI) has played a central role in shaping Japan’s approach to AI governance. Rather than pursuing prescriptive regulation like the EU’s AI Act, Japan has favored a principles-based approach built around guidelines and industry self-regulation — a strategy designed to encourage AI development and adoption while managing risks.
METI’s AI Guidelines, developed in consultation with the AI Strategy Council, outline principles for responsible AI development and use, including transparency, fairness, accountability, and safety. The guidelines address specific concerns relevant to LLMs, including hallucination management, bias mitigation, and the handling of personal information in training data. While not legally binding, these guidelines have become the de facto framework that Japanese enterprises use to evaluate and deploy AI systems.
The Hiroshima AI Process, launched during Japan’s G7 presidency in 2023, extended Japan’s AI governance influence internationally. The resulting Hiroshima Process International Guiding Principles for Organizations Developing Advanced AI Systems and the associated Code of Conduct established a multilateral framework that reflects many of the principles embedded in Japan’s domestic guidelines. This alignment between domestic and international frameworks gives Japanese AI developers a coherent regulatory environment in which to operate.
Compute Infrastructure: The Bottleneck
Training competitive LLMs requires enormous computational resources — specifically, access to thousands of high-end GPUs (or equivalent accelerators) for weeks or months. Japan faces a compute infrastructure gap that constrains its LLM ambitions. While the country has significant supercomputing assets (including Fugaku, which was the world’s fastest supercomputer from 2020 to 2022), access to the GPU clusters specifically optimized for LLM training has been limited.
The government has responded with substantial investment. METI announced a ¥200 billion ($1.5 billion) program to build domestic AI compute infrastructure, including subsidies for organizations establishing GPU clusters in Japan. Several major cloud providers, including Amazon Web Services, Google Cloud, and Microsoft Azure, have announced expansions of their Japanese data center capacity with AI-optimized infrastructure. These investments are essential: without domestically available compute at competitive scale, Japanese LLM developers will be unable to train models that match the capabilities of international competitors.
The Strategic Implications
Japan’s LLM development effort represents more than a technology competition. It is a strategic initiative with implications for the country’s economic competitiveness, national security, and cultural identity in the AI era. The ability to process, reason about, and generate Japanese text at superhuman scale is a capability that will influence everything from government efficiency to corporate productivity to creative expression.
For international businesses, the emergence of Japan’s domestic LLM ecosystem creates both competitive pressure and partnership opportunities. International AI providers must contend with domestic alternatives that better understand Japanese language and business context. At the same time, collaboration opportunities exist — particularly in areas where Japanese and international capabilities are complementary, such as multilingual models, domain-specific fine-tuning, and AI safety research.
The coming years will determine whether Japan’s sovereign AI strategy produces models that are genuinely competitive with international leaders or whether the domestic alternatives remain niche solutions for sovereignty-sensitive applications. The scale of investment, the quality of the research talent involved, and the strong demand signals from Japanese enterprise customers all suggest that Japan’s LLM ecosystem will be a significant force in the global AI landscape.
Interested in Japan’s AI and LLM ecosystem? Contact Japonity — we connect global businesses with Japan’s most innovative companies.
