Rakuten Launches 70B Open Source Japanese LLM

Rakuten, Japan's largest e-commerce company, has officially unveiled an open source large language model featuring 70 billion parameters specifically optimized for the Japanese language. The release marks one of the most significant contributions to the non-English AI ecosystem and positions Rakuten as a serious contender in the global open source AI race alongside Meta, Mistral, and other Western players.

The model represents a major milestone for Japanese-language AI, offering developers and businesses a powerful foundation model that understands the nuances of one of the world's most complex writing systems. Unlike general-purpose multilingual models that often treat Japanese as a secondary language, Rakuten's model places Japanese comprehension and generation at its core.

Key Facts at a Glance

• Model size: 70 billion parameters, placing it among the largest open source Japanese-focused LLMs ever released
• License: Open source, enabling commercial and research use by developers worldwide
• Developer: Rakuten Group, Inc., a $10+ billion Japanese tech conglomerate
• Focus: Native Japanese language understanding, generation, and reasoning
• Architecture: Built on transformer-based architecture with optimizations for Japanese tokenization
• Competition: Directly challenges models from NEC, Preferred Networks, and NTT in the Japanese AI space

Why a Japanese-Specific Model Matters for Global AI

Most leading large language models — including OpenAI's GPT-4, Anthropic's Claude, and Meta's Llama 3 — are primarily trained on English-language data. While these models handle Japanese to varying degrees, they often struggle with the language's 3 distinct writing systems (hiragana, katakana, and kanji), complex honorific structures, and context-dependent grammar.

Rakuten's 70B model addresses these gaps head-on. By dedicating its training pipeline to Japanese-language data, the model achieves significantly better performance on Japanese benchmarks compared to general multilingual models of similar size. This approach mirrors the strategy adopted by companies like France's Mistral AI, which has emphasized European language capabilities alongside English.

The Japanese AI market is projected to reach $25 billion by 2027, according to industry estimates. Having a high-quality open source foundation model could accelerate adoption across sectors ranging from customer service automation to legal document analysis — areas where Japanese language precision is non-negotiable.

Technical Architecture and Training Details

Rakuten's 70B model builds on the transformer architecture that has become the standard for modern LLMs. However, the company introduced several key modifications to better handle Japanese text processing.

One of the most critical innovations involves tokenization. Standard tokenizers used by English-centric models often fragment Japanese text inefficiently, sometimes splitting single characters into multiple tokens. Rakuten developed a custom tokenizer that treats Japanese characters and common phrases more efficiently, reducing the token count needed to represent Japanese text by an estimated 30-40% compared to models like Llama 2's default tokenizer.

The training data reportedly includes a diverse mix of:

• Japanese web content and curated text corpora
• Academic papers and technical documentation in Japanese
• Business and financial documents
• Conversational and informal Japanese text
• Bilingual Japanese-English parallel data for cross-lingual capability

This balanced dataset ensures the model performs well across formal and informal registers — a critical distinction in Japanese, where the difference between casual speech and business language is far more pronounced than in English.

How Rakuten's Model Stacks Up Against Competitors

The Japanese LLM landscape has grown increasingly competitive over the past 18 months. Several major Japanese corporations have entered the race, each taking slightly different approaches to building language models.

NEC released a series of Japanese LLMs with up to 13 billion parameters, targeting enterprise applications. Preferred Networks, a Tokyo-based AI startup valued at over $2 billion, has developed its own foundation models with a focus on scientific computing. NTT, Japan's telecom giant, launched 'tsuzumi,' a lightweight Japanese LLM designed for on-premises deployment.

Rakuten's 70B model dwarfs most of these competitors in raw parameter count. While parameter count alone doesn't determine quality, larger models generally exhibit stronger reasoning capabilities and broader knowledge coverage. Compared to Meta's Llama 3 70B, which serves as a general-purpose multilingual model, Rakuten's offering is expected to outperform on Japanese-specific tasks while potentially trading off some English-language capability.

The open source licensing strategy also differentiates Rakuten from competitors like NTT, which has kept its models more restricted. By making the model freely available, Rakuten aims to build a developer ecosystem around its technology — a playbook borrowed directly from Meta's successful Llama strategy.

What This Means for Developers and Businesses

For developers working on Japanese-language applications, Rakuten's release is a game-changer. Previously, building high-quality Japanese AI applications required either expensive API calls to proprietary models or fine-tuning general-purpose open source models that offered mediocre Japanese performance out of the box.

With a 70B open source model optimized for Japanese, developers can now:

• Deploy locally without relying on external API providers
• Fine-tune the model for domain-specific Japanese applications at lower cost
• Build commercial products without restrictive licensing fees
• Maintain data sovereignty by running inference on private infrastructure
• Reduce latency for Japanese-language services by eliminating API round-trips

For multinational businesses operating in Japan, the model opens doors to more natural customer interactions, better document processing, and improved internal knowledge management. Japan's strict data privacy regulations make locally-deployable open source models particularly attractive for enterprises in finance, healthcare, and government sectors.

Startups in the Japanese market also stand to benefit enormously. The cost barrier to building AI-powered products in Japanese has dropped dramatically with this release, potentially sparking a wave of innovation similar to what Llama's open source release triggered in the English-language AI startup ecosystem.

Rakuten's Strategic Play in the AI Arms Race

Rakuten's decision to invest heavily in AI infrastructure reflects a broader strategic pivot for the company. Known primarily as Japan's answer to Amazon, Rakuten operates an ecosystem spanning e-commerce, fintech, mobile telecommunications, and digital content. AI integration across these verticals could yield massive efficiency gains and new revenue streams.

The company has been investing in AI research for several years, building out its Rakuten Institute of Technology with labs in Tokyo, Singapore, Paris, and other global locations. The 70B model release signals that Rakuten is ready to transition from internal AI research to public-facing AI leadership.

There is also a national dimension to this effort. The Japanese government has identified AI as a strategic priority, allocating billions of yen in subsidies and incentives for domestic AI development. By releasing a world-class open source model, Rakuten positions itself as a national champion in AI — a role that could bring regulatory goodwill and government contracts.

Looking Ahead: The Future of Non-English AI Models

Rakuten's release is part of a growing global trend toward language-specific and region-specific AI models. In 2024 alone, we have seen significant open source releases targeting Arabic, Korean, Chinese, and several European languages. This trend challenges the assumption that a single English-dominant model can serve the entire world.

The implications are profound. As more high-quality non-English models emerge, we may see a fragmentation of the AI landscape along linguistic lines — with different models dominating different language markets. This could create opportunities for regional AI companies while complicating the global ambitions of players like OpenAI and Google.

For Rakuten, the next steps likely include releasing smaller, more efficient variants of the model (such as 7B and 13B versions) for edge deployment, as well as instruction-tuned and chat-optimized versions. The company may also develop specialized models for its core business verticals, including e-commerce product search, financial document analysis, and customer support automation.

The broader AI community will be watching closely to see whether Rakuten's open source strategy translates into a thriving developer ecosystem — and whether other major Asian tech companies follow suit with their own open source releases. One thing is clear: the era of English-only dominance in AI is coming to an end, and Rakuten is helping write the next chapter.