Rakuten releases open-source model Rakuten AI 3.0, configuration file shows based on DeepSeek V3 architecture Odaily Planet Daily News: Lotte Group has released the open-source model Rakuten AI 3.0, calling it "Japan's largest high-performance AI model". This model adopts the MoE architecture, with a total parameter of 671 billion, 37 billion parameters activated for each inference, and a context window of 128K. It is optimized for Japanese scenarios and performs better than GPT-4o in multiple Japanese benchmark tests. This model is one of the achievements of the GENIAC project promoted by the Japanese Ministry of Economy, Trade and Industry and the New Energy Industry Technology Development Agency (NEDO), and has received partial computing power support. Lotte did not disclose the source of the basic model, only stating that it was built based on the achievements of the open source community. The community found in the HuggingFace model file that its config.json contains "deepseek_v3" and related architecture fields, and the parameter size and context configuration are consistent with DeepSeek V3, indicating that the model may be fine tuned in Japanese based on DeepSeek V3.