Rakuten AI 3.0
Rakuten's 671B open-weight Japanese-English model on DeepSeek-V3
Overview
Rakuten AI 3.0 is a 671-billion-parameter mixture-of-experts language model for Japanese and English, built on the DeepSeek-V3 architecture with a 128K context, released as Apache 2.0 open weights on Hugging Face in March 2026. Free to self-host on an 8-GPU node; no public API.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 1 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with Rakuten AI 3.0
- 671B mixture-of-experts Japanese-English model
- Apache 2.0 open weights on Hugging Face
- 128K-token context window
- FP8 weights for SGLang serving
- Rakuten AI Gateway internal deployment
Rakuten AI 3.0 is a large, permissively licensed model tuned for Japanese, which makes it useful for Japanese organisations that want to host their own model. It is a fine-tune of DeepSeek-V3 rather than a model built from scratch — the repository acknowledges DeepSeek's licence — and its benchmark claims are Rakuten's own. Running it needs an eight-GPU server, and there is no hosted API. Worth evaluating for self-hosted Japanese workloads; most teams will find hosted models simpler.
What we found
Rakuten AI 3.0 is a large language model from Rakuten Group, released as open weights on Hugging Face in March 2026 after being unveiled in December 2025. It is a mixture-of-experts model with 671 billion parameters, about 37 billion of them active for each token according to its model card, a 128,000-token context window and support for Japanese and English. Its configuration identifies it as built on the DeepSeek-V3 architecture, and the repository added DeepSeek's MIT licence notice a day after release, following press reports that it had been left out. Rakuten describes the model as leveraging the best of the open-source community and, according to its own research, as the largest LLM disclosed in Japan as of March 2026.
Part of its training cost was covered by GENIAC, the Japanese government's generative-AI programme run by METI and NEDO, and training ran on Rakuten's own multi-node GPU cluster. Rakuten makes it available across its services through the Rakuten AI Gateway, an internal platform, and says trials showed cost reductions of up to 90%; it has not said which customer-facing features use it. Its benchmark results, on Japanese tests such as JamC-QA and M-IFEval against other Japanese models, are Rakuten's own, and we found no independent leaderboard results.
It follows Rakuten AI 7B, which was based on Mistral-7B, and Rakuten AI 2.0, an 8x7B mixture-of-experts model. Rakuten Group, the Japanese internet company behind the model, is based in Tokyo.
The weights are free under the Apache 2.0 licence, with no gating, and stored in FP8. There is no public hosted API, and running it takes serious hardware: Rakuten's own instructions use SGLang across an eight-GPU node.
It is mainly of interest to Japanese organisations that want a large, permissively licensed model they can host themselves.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Rakuten AI 3.0's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Rakuten AI 3.0 unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Rakuten AI 3.0 directly →
Spotted something out of date? Suggest an update →
Alternatives to Rakuten AI 3.0
More in LLMs & Chat
