Rakuten Launches 70B Open-Source Japanese LLM

Rakuten, Japan's largest e-commerce conglomerate, has officially unveiled an open-source Japanese language model boasting 70 billion parameters, marking one of the most significant contributions to non-English AI development in recent memory. The model positions Rakuten as a serious contender in the global large language model race and signals a growing push by Asian tech giants to build sovereign AI capabilities outside the Western ecosystem dominated by OpenAI, Meta, and Google.

The release arrives at a critical moment when governments and corporations worldwide are recognizing that reliance on English-centric models creates both performance gaps and strategic vulnerabilities for non-English-speaking populations.

Key Facts at a Glance

• Model size: 70 billion parameters, making it one of the largest open-source Japanese-focused LLMs ever released
• License: Open-source, enabling researchers and developers to freely access, modify, and deploy the model
• Training data: Built on massive Japanese-language corpora alongside multilingual datasets
• Base architecture: Believed to leverage a transformer-based design similar to Meta's LLaMA family
• Target users: Japanese enterprises, researchers, government agencies, and global developers building Japanese-language applications
• Availability: Released through Rakuten's AI research division with model weights accessible on Hugging Face

Why Rakuten Is Betting Big on Japanese AI

Rakuten's decision to build and open-source a 70 billion parameter model is not merely a technical achievement — it is a strategic business move. The company operates across more than 70 services in Japan, from e-commerce and banking to telecommunications and streaming. Each of these verticals generates enormous volumes of Japanese-language data and customer interactions.

Existing models like GPT-4, Claude, and Gemini perform well in English but often struggle with the nuances of Japanese, a language with 3 writing systems (hiragana, katakana, and kanji), complex honorific structures, and context-dependent meanings. By training a model specifically optimized for Japanese, Rakuten can dramatically improve AI-driven customer service, product recommendations, content generation, and internal automation across its entire ecosystem.

The open-source approach also serves as a talent magnet. By releasing the model publicly, Rakuten attracts contributions from Japan's vibrant research community while establishing itself as a thought leader in Asian AI development.

Technical Architecture and Training Approach

While Rakuten has not disclosed every architectural detail, the 70 billion parameter model is understood to follow the decoder-only transformer architecture that has become the industry standard since the success of GPT-3 and its successors. This puts it in the same parameter class as Meta's LLaMA 2 70B and Alibaba's Qwen-72B, enabling direct performance comparisons.

The training process reportedly involved several key innovations:

• Continual pre-training on Japanese text corpora, building upon a multilingual foundation to retain cross-lingual capabilities
• Instruction tuning using curated Japanese prompt-response pairs to improve conversational and task-following abilities
• Reinforcement learning from human feedback (RLHF) with Japanese-speaking annotators to align outputs with cultural and linguistic expectations
• Efficient tokenization designed specifically for Japanese, reducing the token-per-character ratio that plagues English-centric tokenizers when processing CJK languages

One of the most persistent problems with applying English-first models to Japanese is tokenization inefficiency. Standard BPE tokenizers developed for English often split Japanese characters into multiple tokens, inflating costs and degrading context window utilization. Rakuten's model addresses this head-on with a tokenizer trained on Japanese-heavy data, reportedly achieving 2x to 3x better efficiency compared to using GPT-4's tokenizer on Japanese text.

How It Stacks Up Against Competitors

Rakuten's 70B model enters a crowded but rapidly evolving landscape of Japanese-language LLMs. Several other organizations have released competing models, though none with the combination of scale, openness, and corporate backing that Rakuten brings.

NEC released a series of Japanese LLMs in 2023, ranging from 7B to 13B parameters. CyberAgent, another Japanese tech firm, launched its own open-source models but at smaller scales. Preferred Networks (PFN), a Tokyo-based AI startup, has been developing large-scale Japanese models with government funding. Meanwhile, Stability AI Japan contributed to the Japanese open-source ecosystem before the company's well-publicized financial troubles.

Compared to these efforts, Rakuten's 70B model represents a significant leap in scale. Larger models generally deliver better performance on complex reasoning tasks, nuanced language understanding, and multi-step problem solving — all areas where Japanese-language AI has historically lagged behind English counterparts.

On the global stage, the model competes with multilingual offerings from Meta (LLaMA 3), Alibaba (Qwen 2), and Mistral AI (Mixtral). However, none of these models were specifically optimized for Japanese from the ground up, giving Rakuten a potential quality advantage in Japanese-specific benchmarks like JCommonsenseQA, JNLI, and JSQuAD.

The Strategic Importance of Sovereign AI

Rakuten's release reflects a broader global trend toward sovereign AI — the idea that nations and regions need domestically developed AI capabilities rather than depending entirely on American technology companies. This movement has gained substantial momentum in 2024.

France has backed Mistral AI with billions in investment. The UAE developed Falcon through the Technology Innovation Institute. Saudi Arabia launched ALLaM for Arabic. India is pursuing multiple Hindi-focused LLM initiatives. And now Japan, the world's 4th largest economy, is asserting its AI independence through corporate and government partnerships.

The Japanese government has been particularly active in promoting domestic AI development. The Ministry of Economy, Trade and Industry (METI) has allocated significant funding for AI infrastructure, including compute clusters and training data initiatives. Rakuten's model aligns perfectly with these national priorities.

For Japan specifically, the stakes are high. The country faces a demographic crisis with an aging population and shrinking workforce. AI-powered automation is widely viewed as essential to maintaining economic productivity, making high-quality Japanese-language AI not just a technical goal but a national imperative.

What This Means for Developers and Businesses

The open-source nature of Rakuten's model creates immediate opportunities for a wide range of stakeholders:

• Japanese enterprises can deploy the model on-premises for customer service chatbots, document analysis, and internal knowledge management without sending sensitive data to foreign API providers
• Global companies operating in Japan gain access to a high-quality Japanese AI backbone for localization, translation, and market-specific content generation
• Startups can build Japanese-language AI products without the massive capital expenditure of training their own foundation model
• Researchers can study Japanese NLP at a scale previously only available to well-funded corporate labs
• Government agencies can adopt the model for public-facing services while maintaining data sovereignty

The cost implications are also significant. Running inference on an open-source 70B model using quantized versions (such as GPTQ or GGUF formats) is dramatically cheaper than paying per-token API fees to commercial providers. For high-volume Japanese-language applications, the savings could amount to hundreds of thousands of dollars annually.

Challenges and Limitations Ahead

Despite the excitement, Rakuten's model faces several notable challenges. A 70 billion parameter model requires substantial computational resources to run — typically multiple high-end GPUs like NVIDIA A100s or H100s for full-precision inference. This limits deployment options for smaller organizations without access to significant hardware.

Quantization techniques can reduce resource requirements, but they come with potential quality tradeoffs, particularly for a linguistically complex language like Japanese. Finding the right balance between efficiency and accuracy will be critical for widespread adoption.

There are also questions about the model's performance on specialized domains. General-purpose LLMs often struggle with industry-specific terminology in fields like medicine, law, and finance. Rakuten will need to either provide fine-tuned variants or offer clear guidance for domain adaptation.

Finally, the open-source release raises safety considerations. Without the guardrails of a managed API, the model could potentially be used to generate harmful content in Japanese. Rakuten has indicated it includes safety guidelines and usage policies, but enforcement in the open-source ecosystem remains inherently challenging.

Looking Ahead: Japan's AI Ambitions Grow

Rakuten's 70B model is likely just the beginning. The company has signaled ongoing investment in AI research, and the competitive dynamics in Japan's tech sector virtually guarantee that rivals like SoftBank, NTT, and Sony will accelerate their own LLM efforts in response.

The broader trend is clear: the era of English-only AI dominance is ending. As more nations and companies invest in language-specific models, the global AI landscape is becoming increasingly multipolar. For the estimated 125 million Japanese speakers worldwide, Rakuten's contribution represents a meaningful step toward AI that truly understands their language and culture.

Developers interested in exploring the model can access it through Hugging Face and Rakuten's official AI research channels. With the open-source community already beginning to experiment with fine-tuning and optimization, the ecosystem around this model is expected to grow rapidly in the coming months.