- New “Rakuten AI 3.0” LLM achieves top scores in Japanese language benchmarks
- Delivers up to 90% cost reduction when powering Rakuten Ecosystem services 

Tokyo - Rakuten Group, Inc. today unveiled its latest AI model, Rakuten AI 3.0, a Japanese large language model (LLM) developed as part of the Generative AI Accelerator Challenge (GENIAC) project promoted by the Ministry of Economy, Trade and Industry (METI) and the New Energy and Industrial Technology Development Organization (NEDO).

The new model is based on a Mixture of Experts (MoE)^*1 architecture and is now available for use across Rakuten services through the Rakuten AI Gateway^*2, Rakuten’s internal platform offering a suite of GenAI APIs. Going forward, it will be rolled out sequentially through the Rakuten AI^*3 agentic platform to other Rakuten services. In addition, the model will be released as an open-weight model in Spring 2026. 

Rakuten AI 3.0, the company’s largest LLM^*4 to date, is an approximately 700 billion parameter MoE model optimized for Japanese. Developed by leveraging the best models from the open source community and building on Rakuten’s high-quality, bilingual original data, engineering and research, it offers a superior grasp of Japanese language and culture. Rakuten AI 3.0 balances both performance and efficiency. Extensively trained on Rakuten’s robust data, Rakuten AI 3.0 delivered up to 90% cost reduction in trials when powering Rakuten Ecosystem services, compared with third-party frontier AI models^*5.

In July 2025, Rakuten was selected for the third term of the GENIAC project to develop Japanese language-optimized AI models. Part of the training cost for Rakuten AI 3.0 was provided from the GENIAC project which offers support for computing resources necessary for Japan’s generative AI development.

Ting Cai, Chief AI & Data Officer of Rakuten Group, commented, "At Rakuten, we’re committed to developing high-quality, cost-efficient models that deliver transformative benefits to all – from the small ventures to global enterprises. Our strategy aims to enrich user experiences and provide tangible AI-driven value. With the launch of Rakuten AI 3.0, our most competitive model yet, we are set to elevate our offerings even further."

Watanabe Takuya, Director of AI Industry Strategy Office, Ministry of Economy, Trade and Industry, commented, "We welcome the successful development of large scale, efficient AI models that deliver high performance. We look forward to seeing the implementation of Rakuten’s achievement to expand AI ecosystems and lead Japan’s AI industry.”

High-efficiency model trained at a large scale
The model activates approximately 40 billion parameters for each individual token within the total of approximately 700 billion parameters, while the remaining parameters stay inactive for increased efficiency. The total number of active parameters encompasses three dense layers as well as expert components, where each token routes through eight specialized experts in addition to the always-active shared expert. This innovative structure maintains high efficiency within the model when processing requests.

Model training was carried out on an in-house multi-node GPU cluster engineered by Rakuten. The model is deployed in Rakuten's isolated, secure cloud environment to ensure a high level of control of data security, with all data kept internally. The model is fine-tuned using proprietary, high-quality data optimized specifically for the Japanese market and Rakuten's business needs.

Best-in-class Japanese performance
After further fine-tuning of the foundation models^*6 for conversational and instruction following abilities, Rakuten conducted model evaluations on multi-turn conversational benchmarks also known as Japanese MT-Bench^*7, which assesses model capabilities in actions including adopting and maintaining a persona, composing a compelling blog post or providing insights into technical terms. Scores of the new model were compared with leading models as well as Rakuten AI 2.0^*8 and other previously released Rakuten LLMs.

Rakuten AI 3.0 (LLM) compared to leading models focusing on the Japanese language: 

Rakuten AI LLM comparison

Rakuten is continuously pushing the boundaries of innovation to develop best-in-class LLMs for R&D and deliver best-in-class AI services to its customers. By developing models in-house, Rakuten can build up its knowledge and expertise and create models optimized to support the Rakuten Ecosystem.

*1 The Mixture of Experts model architecture is an AI model architecture where the model is divided into multiple sub models, known as experts. During inference and training, only a subset of the experts is activated and used to process the input.
*2 Rakuten AI Gateway is a unified platform offering a suite of Generative AI APIs tailored for experimentation, development and production needs. Rakuten AI Gateway also allows employees to accelerate both coding and non-coding tasks with advanced AI agents.
*3 Rakuten AI is a cutting-edge agentic platform that serves as a crucial gateway, seamlessly integrating with Rakuten's diverse ecosystem (spanning shopping, fintech, travel, entertainment and communications) to offer personalized customer experiences. Beyond extensive research and agentic capabilities, its strong Japanese contextual processing enables deep integration with local services, aiming to be a trusted companion that improves daily life and society.
*4 According to Rakuten research as of December 18, 2025. Rakuten has also developed Rakuten AI 7B, a 7 billion parameter LLM, and Rakuten AI 2.0, an 8x7B, 47 billion parameter LLM.
*5 According to Rakuten research as of December 18, 2025. When comparing the cost per token (input and output).
*6 Foundation models are models that have been pre-trained on vast amounts of data and can then be fine-tuned for specific tasks or applications.
*7 Results of evaluation tests are carried out on Japanese MT-Bench. Japanese MT-Bench is a set of 80 challenging open-ended questions for evaluating chat assistants on eight dimensions: writing, roleplay, reasoning, math, coding, extraction, STEM, humanities originally introduced for English by Zheng et al. https://arxiv.org/abs/2306.05685  
Rakuten used the standardized implementation from the Institute of Science, Tokyo https://github.com/swallow-llm/swallow-evaluation-instruct/releases/tag/v202510  
Evaluation of responses is conducted with GPT4 (gpt-4o-2024-08-06) as a judge, in line with the public leaderboard, for comparison and transparency in evaluation settings. 
*8 Rakuten AI 2.0, released in February 2025, is based on Rakuten AI 7B, released in March 2024: https://global.rakuten.com/corp/news/press/2025/0212_02.html
The score of the Rakuten AI 2.0 (8x7B-instruct) model in this evaluation is updated due to the change in the judge model version, in line with the public leaderboard.