Rakuten Group is expanding its AI division, now comprising 1,000 employees, and is focusing on cost-efficient AI model development. Led by AI chief Ting Cai, the company utilises thousands of Nvidia chips to support its AI initiatives. Rakuten's AI strategy prioritises minimising costs while augmenting the company's various businesses.
Rakuten AI 3.0, the company's latest large language model (LLM), uses a Mixture of Experts (MoE) architecture with approximately 700 billion parameters, activating only around 40 billion parameters per token. This design lowers computing requirements during inference compared to dense models. Rakuten claims its new model is 90% cheaper to run than comparable LLMs, though this has not been independently verified. The company reported that AI contributed ¥10.5 billion to its operating income in 2024 and aims to double this figure in 2025.
Model training occurred on an in-house multi-node GPU cluster within a secure environment, keeping all data internal. Rakuten AI 3.0 achieved a Japanese MT-Bench score of 8.88, outperforming GPT-4o (8.67) and Rakuten's previous model. The model will be released as an open-weight model in Spring 2026.
