Last updated:
RakutenAI
Rakuten's Japanese-optimised flagship — a ~700B-parameter open-weight model from the GENIAC project.
Model Facts
Latest model: Rakuten AI 3.0
- Provider
- Rakuten
- Released
- Mar 17, 2026
- Context window
- 128K tokens
- Weights
- Open
Full Model List
| Model | Released | Status | Context | Note |
|---|---|---|---|---|
| Rakuten AI 3.0 | Mar 17, 2026 | current | 128K tokens | ~700B-parameter MoE (37B active), 128K context, Apache 2.0 |
- Official model page:
- global.rakuten.com/corp/news/press/2026/0317_01.html
- Official documentation:
- huggingface.co/Rakuten
- Other sources:
- Rakuten AI 3.0 (TelecomTV)
What it is
Rakuten AI 3.0, released on March 17, 2026 as part of the GENIAC project, is a mixture-of-experts model with about 700B parameters (671B total, 37B active per token) and a 128K context window, optimised for Japanese. It is free under the Apache 2.0 license from Rakuten's Hugging Face repository, and Rakuten reports up to 90% cost reduction in trials versus third-party frontier models.
Where this information comes from
Details come from Rakuten's March 2026 press release and launch coverage. Release dates, context windows and benchmark scores are vendor-reported unless stated, and harnesses differ between labs, so treat the numbers as indicative.
Notable traits
- Largest high-performance Japanese model at release.
- Apache 2.0.
- Writing, code generation and document analysis.