Remote OK founder Pieter Levels reignited debate on May 24 over Japan’s AI ambitions, arguing the country lacks a homegrown foundational model and that its flagship offering, Rakuten AI 3.0, is simply a fine-tuned version of China’s DeepSeek. The claim is substantiated: within hours of Rakuten’s March 17 release — billed as “Japan’s largest and most powerful AI model” and backed by Japan’s Ministry of Economy, Trade and Industry (METI) via the GENIAC program — developers on Hugging Face opened the config.json file and found the architecture field explicitly declaring “model_type: deepseek_v3.” The parameter count, a 671B Mixture-of-Experts model with 37B activated per token, matches DeepSeek V3 exactly. Rakuten’s press release described the model as “leveraging the best from the open source community” but did not name DeepSeek.
The fallout was swift. Japanese users criticized the company for receiving government-subsidized compute while not disclosing the base model; Rakuten was also found to have initially removed the original MIT license file from its Hugging Face repository — a compliance violation for any derivative work — before quietly restoring it after community pressure. The episode is part of a broader pattern: Nikkei Asia previously reported that six of the top ten LLMs developed by Japanese companies are secondary builds on either DeepSeek or Alibaba’s Qwen. Japan’s Digital Minister had separately warned public servants against using DeepSeek citing security concerns, making the irony of a METI-funded model built atop it particularly pointed.