Rakuten AI 3.0 Rumored DeepSeek Base

Post LinkedIn

🗾Read original on ITmedia AI+ (日本)

#japanese-llm #deepseek-base #rakuten-strategyrakuten-ai-3.0

💡Rakuten Japanese LLM rumored built on DeepSeek—official response revealed

⚡ 30-Second TL;DR

What Changed

Rakuten released Rakuten AI 3.0 as its latest Japanese LLM

Why It Matters

If confirmed, it highlights Japanese firms' reliance on Chinese open models, potentially impacting data privacy perceptions and competition in regional LLMs.

What To Do Next

Benchmark Rakuten AI 3.0 against DeepSeek-V2 on Japanese NLP tasks like translation.

Who should care:Developers & AI Engineers

🧠 Deep Insight

Web-grounded analysis with 7 cited sources.

🔑 Enhanced Key Takeaways

•Rakuten AI 3.0 is a 671 billion parameter Mixture-of-Experts (MoE) model with 37 billion activated parameters per token and a 128K context length, hosted on Hugging Face under Apache 2.0 license[1][4][6][7].
•The model's config.json file explicitly lists 'model_type': 'deepseek_v3', matching DeepSeek V3's architecture, indicating it is a fine-tuned version rather than independently developed[1][2][3][4][5][6].
•Part of Japan's GENIAC project supported by METI and NEDO, receiving computing power aid for Japanese AI development[6].
•Users observed a pro-China bias in the model's responses on geopolitical questions, contrasting with expected Japanese perspectives[1][2][3].

🛠️ Technical Deep Dive

•Architecture: Mixture-of-Experts (MoE) with 671B total parameters, 37B activated per token[4][6][7].
•Context window: 128K tokens[6][7].
•Config.json specifies 'model_type': 'deepseek_v3', identical to DeepSeek V3 in parameter scale and structure[1][2][4][5][6].
•Supports Japanese and English; quantized versions available in F32, BF16, F8_E4M3 formats on Hugging Face[7].
•Inference recommended via SGLang with specific Docker image and parameters like --tp 8[7].

🔮 Future ImplicationsAI analysis grounded in cited sources

Rakuten AI 3.0 adoption in Japan will grow despite controversy due to superior Japanese benchmarks.

The model outperforms GPT-4o on Japanese culture, history, and instruction-following tasks, positioning it as a leading local option amid government-backed GENIAC support[1][2][3][6].

Increased scrutiny on open-source model transparency will emerge from this incident.

Explicit config.json evidence of DeepSeek V3 base challenges claims of independent development, highlighting risks in rebranding fine-tuned foreign models[1][4][5].

⏳ Timeline

2026-03

Rakuten launches Rakuten AI 3.0 as open-source Japanese MoE model on Hugging Face

2026-03-17

Announcement of Rakuten AI 3.0 via PANews, claiming superiority over GPT-4o in Japanese benchmarks

📎 Sources (7)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

🗾Read original article on ITmedia AI+ (日本)

📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #japanese-llm

Same product