What Is Rakuten AI 3.0? A Clear and Thorough Guide to Rakuten’s Latest LLM Advancing Practical Japanese AI

On March 17, 2026, Rakuten Group announced the launch of “Rakuten AI 3.0,” a high-performance AI model described as one of the largest of its kind in Japan. Following its initial development announcement in December 2025 and subsequent improvements, it has now moved to full public release at no cost. As a large language model optimized for Japanese, it is designed with strong emphasis on use cases such as writing, code generation, document analysis, and information extraction, and it is drawing attention as a model that may push the use of generative AI by Japanese companies into a far more practical stage.

This article will be especially useful for readers such as corporate IT staff who want to introduce a generative AI with strong Japanese-language capability into business operations, DX teams considering an in-house AI foundation, legal, HR, and general affairs departments that handle large volumes of Japanese documents and domestic regulations, developers looking for an open, Japan-oriented model that can be embedded into custom applications, and business professionals who want to clearly understand the difference between “a service like ChatGPT” and “an AI model a company can operate itself.” Rather than simply following the topic because it is in the news, this article is organized from a practical perspective for those who want to judge whether it is truly usable.

To state the conclusion first, the value of Rakuten AI 3.0 lies in the fact that Rakuten has publicly released a large-scale model strong in Japanese under a relatively easy-to-handle license. According to Rakuten’s official announcement, this model is a Mixture of Experts (MoE) model with approximately 700 billion parameters developed as part of the GENIAC project, and it achieved high scores on major Japanese-language benchmarks. In addition, because it is offered under the Apache 2.0 license, companies and developers can more easily evaluate, improve, and integrate it. In other words, Rakuten AI 3.0 is best viewed not simply as “Rakuten’s internal AI,” but as a publicly available foundation model intended for real-world Japanese-language use.

The Background of Rakuten AI 3.0: Why Is Rakuten Releasing a Large-Scale Model Now?

To understand Rakuten AI 3.0, it is essential to look at “GENIAC” (Generative AI Accelerator Challenge), promoted by Japan’s Ministry of Economy, Trade and Industry and NEDO. GENIAC is a national project aimed at strengthening Japan’s ability to develop generative AI foundation models by supporting access to computing resources and development assistance. Rakuten has been developing its model within this framework. What this shows is that Rakuten AI 3.0 is not an isolated corporate experiment, but a project born within the broader movement to strengthen Japan’s domestic AI foundation.

Across the Rakuten Group as a whole, AI is not treated as a one-off new business, but as a central pillar of company-wide strategy. Rakuten promotes the concept of “AI-nization,” making clear its intention to spread AI across a wide range of services including shopping, finance, travel, and entertainment. For a company with operations in 30 countries and regions, more than 70 services, and over 2 billion service users, AI serves as a common layer for creating new value. It is easiest to understand Rakuten AI 3.0 as one of the foundations supporting that broader vision.

From a practical perspective, this background also reflects issues that are especially common among Japanese companies. Many firms want to use generative AI, but face problems such as “the Japanese sounds unnatural,” “it is difficult to handle internal documents,” “we want to reduce dependence on external APIs,” and “it is hard to balance cost and governance.” Rakuten AI 3.0 is a model that appears to offer one answer to these concerns from both sides: Japanese-language performance and openness. It seems aimed at filling operational gaps that are not easy to address simply by using cutting-edge overseas models as they are.

Basic Specifications of Rakuten AI 3.0: The Key Points to Understand First

In Rakuten’s official announcement on March 17, 2026, Rakuten AI 3.0 is described as a model using a “Japanese-optimized Mixture of Experts (MoE) architecture with approximately 700 billion parameters.” MoE is a design in which not all weights are fully activated at all times; instead, only the “experts” needed for each input are selected and used. As a result, it is known as a structure that can balance performance and computational efficiency even at very large scale. In the December 2025 development announcement, the model was described as having roughly 700B total parameters and about 40B active parameters.

Meanwhile, the model card published on Hugging Face states that the total parameters are 671B, the activated parameters per token are 37B, and the context length is 128K. This is a point worth looking at carefully. It is reasonable to interpret Rakuten’s official PR as giving rounded figures like “about 700 billion” and “about 40 billion,” while the model card gives more exact figures such as “671B” and “37B.” When referring to these numbers in articles or internal explanations, it is helpful to avoid misunderstanding by stating both: “Rakuten’s official announcement says approximately 700 billion, while the public model card states 671B.”

The supported languages are Japanese and English. However, the core of its value clearly lies in its Japanese optimization. Rakuten explains that the model was developed using its own high-quality bilingual data, technical capabilities, and research成果, and it highlights understanding of Japanese-specific nuances, culture, and conventions as a strength. From the perspective of enterprise use, what matters is not merely whether the model can respond in Japanese, but whether it is suitable for the kinds of heavy but often unglamorous Japanese-language tasks that occur in real organizations, such as internal approvals, contract assistance, FAQ creation, summarization, extraction, and classification. In that sense, Rakuten AI 3.0 can be seen as a model with a design philosophy closely aligned with the realities of Japanese workplaces.

What Makes It Impressive? Capabilities Revealed Through Japanese Benchmarks

The biggest reason Rakuten AI 3.0 is attracting attention is the benchmark results Rakuten has published. In the announcement on March 17, 2026, the company presented comparisons across several metrics including JamC-QA, MMLU-ProX (Japanese), MATH-100 (Japanese), and M-IFEval (Japanese). Rakuten AI 3.0 scored 76.9 on JamC-QA, 71.7 on MMLU-ProX (Japanese), 86.9 on MATH-100 (Japanese), and 72.1 on M-IFEval (Japanese). As a comparison point, the same announcement listed gpt-4o as scoring 74.7, 64.9, 75.8, and 67.3, respectively.

The important thing in reading these numbers is not to interpret them as meaning “universally the strongest in the world.” Benchmarks are comparisons under specific evaluation conditions, and in real-world operations many factors matter, such as safety controls, response speed, inference cost, tool integration, integration with internal data, and long-context stability. Even so, it is meaningful that the model showed high-level results across multiple indicators involving Japanese knowledge, reasoning, mathematics, and instruction following. This is especially relevant because strengths in these areas are often directly tied to ease of use in practical tasks such as Japanese business document handling and tasks that require fine-grained compliance with instructions.

In the December 2025 announcement, Rakuten also reported a score of 8.88 on the Japanese version of MT-Bench, exceeding gpt-4o’s 8.67. In addition, the company highlighted improvements over Rakuten AI 2.0 and Rakuten AI 7B. What this suggests is that Rakuten AI 3.0 is not merely a larger successor model, but one that has been refined with strong attention to dialogue quality and instruction-following ability. In internal business use especially, it is not enough for a model to simply produce natural Japanese; it also needs to stay aligned with user intent, show less variation in expression, and format outputs according to business instructions. Strengthening those qualities has major practical significance.

For example, even with the same instruction, such as “Summarize the sales meeting minutes and divide them into decisions made, pending items, and tasks by responsible person,” a model with weak Japanese optimization may produce inconsistent heading granularity, fail to separate responsibilities clearly, or make unnatural choices in honorific style. The value of Rakuten AI 3.0 lies in how far it can reduce those kinds of subtle awkwardness and approach a level of quality where its output can be used directly as a business draft. Benchmarks serve as an entry point showing that possibility, and in that sense they are quite meaningful.

What Kinds of Use Cases Is It Good For? Practical Scenarios by Business Function

According to Rakuten’s official explanation, Rakuten AI 3.0 excels in a wide range of text-processing tasks such as writing, code generation, document analysis, and extraction. When interpreted in the context of actual workplaces, the range of possible use cases is quite broad. It appears especially well suited to work that sits between fixed routine tasks and fully unstructured tasks. In other words, it is a strong fit for “semi-structured” work that is hard to automate completely, but where preparation and draft generation can make a large difference.

For example, in public relations or marketing departments, likely use cases include first drafts of product descriptions, ad copy variations, FAQ drafts, campaign announcements, review summaries, and initial competitor comparison tables. In HR, it could help with formatting job descriptions, summarizing interview notes, drafting internal policy Q&A, and restructuring training materials. In legal and general affairs, it may be useful for extracting clauses from contracts, organizing issues in internal regulations, summarizing differences across multiple documents, and classifying inquiries in the first stage. In development teams, realistic uses include code generation and completion, generating sample code from specifications, summarizing key points from logs, and writing explanations for existing functions.

Let us look at one specific example. Suppose an e-commerce company wants to create response templates for a customer-facing chatbot based on three internal documents: a return policy, delivery conditions, and member benefits. A Japanese-strong model like Rakuten AI 3.0 is likely to be a very good fit here. If the company provides those internal rule documents as input and specifies conditions such as “use polite style,” “avoid overly definitive statements,” “list exceptional conditions in bullet points,” and “if the basis is unclear, stop and request confirmation,” it becomes easier to balance FAQ generation accuracy with readability. This style of use is not based on the idea of “letting generative AI do everything,” but on the idea of “quickly creating drafts that humans can verify more easily.” In enterprise deployment, that mindset is extremely important.

It may also have value in settings such as local governments, educational institutions, administrative departments around hospitals, and professional services offices, where large volumes of Japanese documents are handled and both politeness and accountability matter. That said, in high-risk domains such as medical judgment or legal judgment, it would be inappropriate to operate by accepting the model’s output as-is. Rakuten AI 3.0 is a powerful foundation, but it is safer to view it not as a tool for replacing human judgment, but as a foundation for organizing information and supporting draft creation.

The Meaning of Open Release: What Changes for Companies?

What makes Rakuten AI 3.0 even more valuable is that it is offered free of charge under the Apache 2.0 license. For companies, this has very practical implications. It provides a high degree of freedom for commercial use, modification, and redistribution, making it easier to consider integration into internal systems and additional training. Of course, in actual use it is still necessary to review the license text and dependencies, but at the very least it represents a major step forward from the situation where “it is unclear whether we are even allowed to use it, so we cannot begin evaluation.”

When relying only on general cloud AI services, companies must always worry about API fees, data transmission destinations, log retention, the impact of changes in terms of use, and model replacements. By contrast, open models offer more room for operation in a company’s own environment, domestic cloud infrastructure, or secure dedicated foundations. This difference can directly determine whether adoption is possible, especially for departments that handle documents that are difficult to send outside the organization. Rakuten AI 3.0 can be seen as a model that directly addresses the needs of companies wanting to “use a high-performance Japanese LLM while evaluating and controlling it in-house.”

For example, in sectors such as finance, insurance, telecommunications, public services, and large-scale manufacturing, the barrier to introducing generative AI is often not performance itself, but governance and accountability. If a model exists only behind an external API, progress can become difficult due to security reviews and audit requirements. With an open model, by contrast, it becomes easier to design systems tailored to internal requirements, including internal search, RAG, permission control, audit logs, banned-word filtering, and controlled output templates. Rakuten AI 3.0 attracts attention not only because of its technical strength, but because it increases the freedom of enterprise implementation.

But Do Not Overestimate It: Important Cautions Before Adoption

One very important point is that even though Rakuten AI 3.0 is an excellent model, it is not free from the fundamental limitations of generative AI. The public model card also clearly states that bias, inaccuracy, and safety issues may arise, and that sufficient caution and guardrails are necessary in operation. This is not unique to Rakuten AI 3.0; it is a property shared by large language models in general. In other words, high performance and being always correct are two different things.

There are three especially important points to watch in enterprise use. First, factual error. Models can make mistakes in the most natural Japanese and with the greatest apparent confidence, so human verification is essential in areas where accuracy matters, such as contracts, medicine, accounting, regulation, investor relations, and journalism. Second, information leakage. Even with an open model, problems can arise in handling confidential information if the design of input data handling, inference environments, or log storage is inadequate. Third, do not assume that prompt engineering alone can fully raise quality. In real practice, input preprocessing, rule-based control, limiting reference documents, fixed output formats, and review workflow design are all crucial.

For example, in internal use, instead of giving a vague prompt such as “Extract only the important issues from these meeting minutes that matter for management decisions,” it is more stable to impose constraints such as “Limit issues to three or fewer, quote supporting text from the original, and do not write speculation.” In addition, simply using RAG to limit the reference sources and requiring the output to include the source document name at the end can significantly improve practical usability. The more powerful the model, such as Rakuten AI 3.0, the more results come from “designing the operation” rather than “leaving it to the model’s intelligence.”

How Is It Different from Other AI? Comparison Points Users Should Understand

One point users may easily confuse is that “Rakuten AI” and “Rakuten AI 3.0” can look like they mean the same thing. The former may be used in the context of Rakuten’s overall AI strategy or AI experiences used across services, including agent-like functionality. Rakuten AI 3.0, by contrast, refers specifically to the foundation large language model within that broader context. Put another way, Rakuten AI 3.0 is closer to the “engine” supporting conversational services and business apps than to “the conversational service itself.”

It is also necessary to separate strengths and weaknesses when comparing it with polished SaaS-style AI such as ChatGPT. SaaS-style offerings are often quick to adopt, have well-developed user interfaces, and may include mature support for voice, images, and tool integration. On the other hand, the model itself is often a black box, and fine-grained control or custom modification can be difficult. Open models like Rakuten AI 3.0 are the opposite: they require more effort in operational design and environment setup, but their attraction lies in easier adaptation to Japanese business needs and company-specific requirements. It is better not to ask which is superior in general, but which is better suited to a given purpose.

For example, a company that wants to distribute generative AI to all departments within a few days will likely be better served by a SaaS-style product first. By contrast, if the priority is internal document search, department-specific assistants, integration with custom workflows, on-premises-like operation, or fine-tuning specialized for domestic business use, then the value of an open model like Rakuten AI 3.0 becomes much greater. If adoption begins without clarifying this comparison, misunderstandings are likely, such as “it is not as easy to use as we expected” or “it is not as flexible as we thought.”

Who Should Consider Rakuten AI 3.0?

The organizations likely to benefit most are those that handle large amounts of Japanese documents and want to integrate AI deeply into their business processes. Specifically, this includes mid-sized to large enterprises looking to improve internal knowledge search, e-commerce, telecom, and financial firms seeking better FAQ generation and first-line inquiry handling, back-office teams wanting to reduce the burden of organizing legal, HR, and general affairs documents, and software companies aiming to build custom AI applications. It is likely to be an especially compelling option for workplaces that have felt that English-centered models are still “not quite right” for Japanese work.

On the other hand, it is not necessarily the best immediate fit for every user. For individuals who simply want casual conversation, or who want to immediately use image generation, browsing, and voice interaction together, consumer-ready services may often be easier to understand and use. Precisely because Rakuten AI 3.0 is valuable as a model itself, it requires a certain level of understanding and organizational readiness from the adopting side. In other words, it is less for people looking for a convenient app, and more for those who want to use AI as infrastructure.

One useful guideline is whether an organization can identify at least ten internal pain points that generative AI could improve. These might include organizing meeting minutes, summarizing inquiries, integrating product information, formatting specifications, assisting with code, searching regulations, and drafting reports. If multiple such business issues are already visible, then evaluating Rakuten AI 3.0 as a core foundation is certainly worthwhile. By contrast, adopting it only because it is a hot topic can lead to overly high expectations and poor long-term use.

What to Watch Going Forward: Where Is the Competitive Focus of Japanese AI Headed?

What the arrival of Rakuten AI 3.0 suggests is that competition in Japanese AI is shifting from the stage of “simply building a larger model” to the stage of “whether it is easy to implement in enterprises” and “whether it is truly useful in Japanese workplace settings.” The key things to watch next are not just benchmark updates. Compatibility with RAG, long-context stability, reproducibility in internal document processing, ease of additional training, inference cost, inference environment readiness, and integration with agent-based systems are all likely to shape practical evaluation.

Another important point is the potential spread of the model as an open foundation. Once a model is publicly released, external developers and companies can share evaluation results, and derivative uses such as inference optimization, distillation, continued training, and industry-specific tuning become more likely. At that point, Rakuten AI 3.0 begins to carry value not only as Rakuten’s own achievement, but as common infrastructure for Japanese AI development. In that sense as well, this public release is meaningful because it can enrich the Japanese LLM ecosystem.

Personally, I feel that the real evaluation of Rakuten AI 3.0 will not be determined by “impressive specifications,” but by “how much it improves working time and quality in the workplaces that adopt it.” For example, whether it can cut the pre-review stage of legal work in half, significantly reduce the effort required to create FAQs, raise the self-service resolution rate for internal inquiries, or reduce routine work for development teams. Only when such grounded results accumulate does the value of Japanese LLMs become truly real. Rakuten AI 3.0 is a model that suggests that possibility at a very high level.

Summary

Rakuten AI 3.0 is a Japanese-optimized large language model developed by Rakuten as part of the GENIAC project and launched on March 17, 2026. According to the official announcement, it is an MoE model with approximately 700 billion parameters, it achieved high scores on Japanese benchmarks, and it was publicly released free of charge under the Apache 2.0 license. This is a major step forward for companies and developers who have been seeking a strong, open Japanese-language LLM.

What is especially important is that its value lies not simply in “being domestic,” but in whether it is easy to use for Japanese business work, easy to implement according to company-specific requirements, and possible to operate under proper control. Rather than competing directly with SaaS-style AI, it is better understood as a foundation for internal enterprise use and custom application development in Japanese companies. For workplaces aiming to balance Japanese-language quality and implementation freedom in use cases such as writing, code generation, document analysis, and extraction, it looks likely to become one of the leading candidates.

At the same time, as a generative AI model, it still carries issues such as error, bias, and safety concerns. In deployment, outcomes will depend not only on the model’s raw capability, but also on operational preparation such as limiting reference information, setting guardrails, designing review processes, and audit design. Rakuten AI 3.0 is not a magical box, but it is a highly realistic and powerful foundation for actually using Japanese AI in business. For those who want to seriously compare options in Japanese generative AI, it is fair to say that this is a model worth checking right now.

Reference Links

• Rakuten Group, Inc.: Press Release, March 17, 2026, “Launch of Rakuten AI 3.0”
• Rakuten Group, Inc.: Press Release, December 18, 2025, “Development Announcement of Rakuten AI 3.0”
• Rakuten / RakutenAI-3.0 (Hugging Face)
• Rakuten Group, Inc.: Official Rakuten AI Page
• METI: Overview of GENIAC (Generative AI Accelerator Challenge)
• NEDO: GENIAC-Related Announcement