Rakuten launches open-weight LLM, claims beat GPT-4o

Rakuten Group has unveiled Rakuten AI 3.0, a Japanese large language model with around 700 billion parameters.

It was developed as part of the government-supported Generative AI Accelerator Challenge project in Japan.

The model uses a mixture of experts architecture and is now available across Rakuten’s services via its internal AI platform.

Rakuten plans to release it as an open-weight model in spring 2026, or around March-May.

Rakuten AI 3.0 achieved the highest score, 8.88, on the Japanese MT-Bench conversational benchmark, ahead of GPT-4o and other Japanese-focused LLMs.

Internal trials showed up to 90% cost savings for Rakuten services compared to third-party AI models.

🔗 Source: Rakuten

🧠 Food for thought

Implications, context, and why it matters.

Rakuten AI 3.0 scored 8.88 on Japanese MT-Bench. The test uses 80 multi-turn Q&A prompts and judges with GPT-4o, which can bias results 1.
Chatbot Arena is an open crowdsourced platform. It turns thousands of pairwise human votes into a ranking with Bradley-Terry modeling (a statistical method that turns pairwise comparisons into a global ranking) 2.
Rakuten AI 2.0 scored 7.08 on Japanese MT-Bench in February 2025 1. The rise to 8.88 marks progress. Claims of beating GPT-4o need checks on Japanese Chatbot Arena, the Japanese-language version of the crowdsourced leaderboard 3.
Rakuten self-reported the 90% cost savings. It likely mirrors its own workloads. Examples include prompt mix, latency targets, and hardware, not broad performance.

Rakuten plans an open-weight release in spring 2026, which means public access to parameters and matches prior releases 4. It could allow commercial use and fine-tuning with no license fees, subject to the final license.
Cloud providers and hosting platforms can prepare to support deployment of the 700 billion-parameter model. Serving will need high-memory GPUs and routing, given its mixture of experts architecture (an approach that activates a subset of expert subnetworks per request).
Enterprise software vendors in Japan can build domain fine-tuning services. That means customizing the base model with proprietary data, and Rakuten’s earlier models supported use as bases for other models 4.
The months before release give compliance tooling vendors time to build Japanese-language safety and alignment capabilities (mechanisms to keep model behavior within desired guidelines) tailored to the model’s architecture.