Rakuten Group CEO Hiroshi Mikitani unveiled Rakuten AI 3.0 with much fanfare, positioning it as "Japan's largest high-performance AI model" and releasing it under Apache 2.0 license. The announcement claimed it outperformed GPT-4o on multiple Japanese benchmarks. But within hours, the community spotted something awkward on the HuggingFace page: the config.json explicitly stated model_type: deepseek_v3 and architectures: DeepseekV3ForCausalLM.
The parameter details — 671B total parameters, 37B activated, 128K context window — all match DeepSeek V3 exactly. Rakuten vaguely mentioned "fully leveraging the best results from the open-source community" without naming DeepSeek. While using DeepSeek V3 under Apache 2.0 is legal, the lack of transparency has drawn criticism.
Government-funded project rides on Chinese base model
Notably, Rakuten AI 3.0 is an output of the GENIAC program (Generative AI Accelerator Challenge), jointly promoted by Japan's Ministry of Economy, Trade and Industry (METI) and the New Energy and Industrial Technology Development Organization (NEDO). The Japanese government provided partial computing subsidies. Rakuten used those subsidies to finetune a Chinese open-source model and rebranded it as its own — a fact that raises eyebrows both domestically and internationally.
DeepSeek V3: 20x cheaper training, same benchmark punch
DeepSeek V3 launched in December 2024, trained at a cost of roughly $5-6 million — over 20 times cheaper than GPT-4 — while matching or beating top closed-source models on multiple benchmarks. Its Apache 2.0 license allows anyone to commercialize and republish, which is precisely what Rakuten did. The irony: DeepSeek hasn't spent a dime on marketing, yet it has become the most effective ad in the open-source AI world.