RakutenAI-3.0.1

Model Description

RakutenAI-3.0.1 is a post-trained version of RakutenAI-3.0. RakutenAI-3.0 is an approximately 700 billion parameter Mixture of Experts (MoE) model optimized for Japanese. Developed by leveraging the best from the open-source community and building on Rakuten’s high-quality, bilingual original data, engineering and research, it offers a good grasp of Japanese language and culture. RakutenAI-3.0 was developed through continual pre-training on top of the DeepSeek-V3-0324 base model which completed a security evaluation in a controlled, in-house sandbox environment.

Compared to DeepSeek-V3-0324 and RakutenAI-3.0, RakutenAI-3.0.1 shows stronger alignment with broad safety principles.

Single-turn harmful request evaluation results on Japanese and English, the numbers indicate the harmless response rate - higher is better.

For more details, please check the blog at:

• ENGLISH Version
• JAPANESE Version

Model Details

• Developed by: Rakuten Group, Inc.
• Language(s): Japanese, English
• License: Apache License, Version 2.0
• Model Architecture: Mixture of Experts (MoE)
• Base model: RakutenAI-3.0
• Total Parameters: 671B
• Activated Parameters per Token: 37B
• Context Length: 128K

How to Run Locally

Inference with SGLang

• Recommended DOCKER Image: lmsysorg/sglang:v0.5.13

python -m sglang.launch_server \
    --model-path Rakuten/RakutenAI-3.0.1 \
    --tool-call-parser deepseekv3 \
    --tp 8

For additional inference and deployment options, see the SGLang cookbook.

Limitations and Bias

RakutenAI-3.0.1 can generate human-like text across a wide range of topics. However, like all large language models, it has limitations and may produce biased, inaccurate, unsafe outputs, or outputs that may infringe the rights of third parties. Please exercise caution and judgment when interacting with this model, and ensure appropriate safeguards are in place for production deployments.

Disclaimer regarding the European Union (EU) Market

This model is NOT intended for, and is NOT made available to, users or entities residing within the European Union (EU) or the European Economic Area (EEA). The Provider has not placed, and does not intend to place, this model on the EU market.

Dataset Attribution

This project uses portions of the safety-related datasets listed below for model fine-tuning and evaluation, among other annotated or synthetic data. The datasets, or portions thereof, may have been processed, transformed, or used as source material for synthetic data generation.

• AdvBench — MIT License
• CatQA — Apache License 2.0
• DiaSafety — Apache License 2.0
• Gandalf — MIT License
• GEST — Apache License 2.0
• IPBench — Open Data Commons Attribution (ODC-By)
• JADE — MIT License
• SafeText — MIT License
• SGXSTest — Apache License 2.0
• WMDP — MIT License
• Wikipedia — CC-BY-SA
• XSafety — Apache License 2.0

Base Model Attribution

RakutenAI-3.0.1 is a post-trained version of RakutenAI-3.0 built upon DeepSeek-V3-0324 which was released under the MIT License. We acknowledge the authors and comply with their respective licensing terms.

Citation

@misc{rakutengroup2026rakutenai3.0.1,
  author = {Rakuten Group, Inc.},
  title = {RakutenAI-3.0.1},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/Rakuten},
}

Contact

For questions or feedback, please open an issue on this repository or visit ai.rakuten.com.