RakutenAI 2.0

RakutenAI 2.0 was unveiled by Rakuten Group in February 2025 as the company's first Japanese large language model based on a mixture-of-experts architecture. The 8x7B foundation model derives from RakutenAI-7B, routes each token through two of eight experts (13 billion active parameters), and was continually trained on high-quality Japanese and English data.\n\nReleased to the open-source community under Apache 2.0 with an instruct-tuned variant, it achieved leading Japanese MT-Bench scores among open Japanese models at release. A 1.5B mini variant was released alongside for small-device deployments.