A digital representation of high-density neural network pathways emphasizing localized data processing nodes.
TOKYO · March 24, 2026 : Rakuten Group has unveiled Rakuten AI 3.0, a 700-billion parameter model that signals a decisive shift toward sovereign, localized AI. Developed under the Ministry of Economy, Trade and Industry’s (METI) GENIAC project, the model is not merely a play for scale; it is a targeted strike at the linguistic and cultural gaps found in Western-centric frontier models. By prioritizing local "context" over generic data, Rakuten is establishing a defensive moat rooted in cultural nuance.
Strategic Efficiency via MoE
The model utilizes a sophisticated Mixture of Experts (MoE) architecture. While the total parameter count sits at 700 billion, the system activates only 40 billion parameters: roughly 6% of its capacity: per token. This design facilitates a 90% reduction in operational costs compared to third-party frontier models in internal trials. High performance is maintained through an 8-expert configuration that specializes in specific tasks like data extraction and document analysis. This efficiency is critical as agentic AI moves into retail operations, where low-latency and cost-effective reasoning are required at scale.
The Honorific Moat
While global models often struggle with the rigid hierarchies of Japanese honorifics (keigo), Rakuten AI 3.0 treats cultural context as core business logic. Trained on Rakuten’s proprietary bilingual data, the model achieved a Japanese MT-Bench score of 8.88, surpassing GPT-4o’s 8.67. For industries like hospitality and retail, this linguistic precision is not a luxury; it is a functional requirement for maintaining brand trust. By releasing the model under an Apache 2.0 license, Rakuten is anchoring itself as the infrastructure for domestic AI, mirroring Seoul’s push for sovereign AI and reinforcing the trend of jurisdictional data control.
Source: https://global.rakuten.com/corp/news/press/2026/0317_01.html
Tags: Tokyo · Consumer Brands / Tech / AI · bcdW Current Today : March 24, 2026
