Rakuten Ships RakutenAI-3.0 but Loses the Story to a Config File
Rakuten open-sourced a 671B-parameter Mixture-of-Experts model tuned for Japanese, then got caught reusing DeepSeek-V3 weights without proper attribution.
ProTakeaways
- Rakuten released RakutenAI-3.0, a 671B MoE model with 37B active parameters and 128K context.
- Available free on Hugging Face under Apache 2.0, optimized for Japanese language and culture.
- Developed under Japan's GENIAC project, backed by METI and NEDO government funding.
- Rakuten claims it beats GPT-4o on Japanese benchmarks including culture, history, and instruction following.
- Community found the config lists 'deepseek_v3' as model type, matching DeepSeek-V3 parameter counts exactly.
- Runs via SGLang with 8-GPU tensor parallelism; quantizations already available for llama.cpp and Ollama.
Rakuten has pushed its largest language model to date onto Hugging Face under an Apache 2.0 license, calling it the biggest high-performance AI model developed in Japan. RakutenAI-3.0 is a 671-billion-parameter Mixture of Experts model with 37B activated parameters per token and a 128K context window, tuned specifically for Japanese language and culture. Within hours of the release, developers digging through the repo config file surfaced evidence that the model is a fine-tune of DeepSeek-V3 rather than a from-scratch build, triggering a licensing controversy that has become as newsworthy as the release itself.
What the model actually is
The model was developed as part of the Generative AI Accelerator Challenge (GENIAC), promoted by Japan's Ministry of Economy, Trade and Industry (METI) and the New Energy and Industrial Technology Development Organization (NEDO). Rakuten positions it as Japan's largest high-performance AI model for companies and professionals building AI applications, with part of the training cost covered by GENIAC, which offers compute resources for Japan's generative AI development.
The Hugging Face model card lays out the specs cleanly:
- Architecture: Mixture of Experts (only a subset of experts activates per token)
- Total parameters: 671B
- Activated parameters per token: 37B
- Context length: 128K tokens
- Languages: Japanese and English
- License: Apache 2.0, free download
- Weights: BF16 and FP8 (E4M3) tensors available
Rakuten says RakutenAI-3.0 is optimized for Japanese and excels at writing, code generation, and document analysis and extraction, with significantly higher accuracy and more robust performance than earlier releases in the family. Those earlier entries were much smaller: Rakuten AI 7B at roughly 7 billion parameters, and Rakuten AI 2.0 with an 8x7B (47B) configuration.