Just a few days ago, Rakuten Group (RakutenThey grandly announced what they claimed was “Japan’s largest domestically developed high-performance AI model”: Rakuten AI 3.0. The model has approximately 700 billion parameters, adopts a Mixture of Experts (MoE) architecture, and was released as open source under Apache 2.0 license. In Rakuten’s official press release, it was described as a milestone for Japan’s AI industry. However, this “Made in Japan” milestone, less than a day later, was caught by netizens from one line of configuration code in their own open source repository that it was actually using Chinese DeepSeek V3 reskin
Less than a day after the “Japan’s Strongest” announcement, the underlying code was completely exposed
Rakuten AI 3.0’s model weights were uploaded to open-source platforms on the day of release. Hugging Face The official page. DeveloperdiscoveredIt’s written in black and white in the configuration file:"model_type": "deepseek_v3"and "architectures": ["DeepseekV3ForCausalLM"]The hidden size of 7168, the number of hidden layers of 61, and all other parameters perfectly match the original DeepSeek V3 settings. In other words, Rakuten’s claimed “approximately 700 billion parameters” is nothing more than a repackaged version of DeepSeek V3’s original 671 billion parameters, simply rounded up.
Delete MIT license, add NOTICE file: ‘License laundering’ controversy erupts
DeepSeek V3 adopts the MIT open-source license, allowing free commercial use and modification, with the only requirement being to retain the original author’s copyright notice. However, Rakuten initially uploaded Rakuten AI 3.0 Until Hugging Face directly deleted DeepSeek’s MIT license documentation and replaced it with their own Apache 2.0 license, repackaging it as a “free open source” release—a practice the open source community criticized as “license laundering” (License Laundering).
Rakuten’s calculation was clear: erase DeepSeek’s name, slap on their own Apache 2.0 license, and repackage themselves as a Japanese AI with a “generously open-sourced 700-billion parameter large model.” After public backlash grew, Rakuten hastily added a file named NOTICE to the open-source repository, acknowledging DeepSeek’s copyright ownership—a move described as “a belated fix with an insincere attitude.”
Took Japanese taxpayers’ money but didn’t mention DeepSeek at all
Rakuten AI 3.0 is an outcome 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). Rakuten was selected in July 2025 and received partial training compute subsidies from GENIAC. However, in its press release, Rakuten only mentioned “leveraging the best from the open source community” without naming DeepSeek at all.
Japanese AI researcherうみゆき（umiyuki）They vented directly on X: “Last time (referring to when Rakuten released RakutenAI-7B), they honestly stated that it was ‘developed based on Mistral-7B’—so why did they suddenly become vague about it this time?” Japanese netizens were even more outraged, pointing out: “They received subsidies from the Japanese government, yet fine-tuned using China’s DeepSeek—this is emotionally unacceptable.”
Won the benchmark, but the opponent was only a 120 billion parameter model
Rakuten officially announced that Rakuten AI 3.0 “surpasses GPT-4o” in multiple Japanese benchmark tests. However, a closer look at the comparison targets reveals that the two models Rakuten benchmarked against are GPT-OSS-Swallow-120B-RL-v0.1 (120 billion parameters) and ABEJA-QwQ32b-Reasoning-Japanese-v1.0 (32 billion parameters). Using a 671 billion parameter model to compare against models with 120 billion and 32 billion parameters gives it a massive scale advantage—yet it completely avoids going head-to-head with models of comparable size.
DeepSeek V3: The “AI Black Ship” Shaking the World
DeepSeek V3, released in December 2024, achieved training costs of only approximately 5 to 6 million USD (approximately NT$195 million to NT$234 million), yet matched or even surpassed top-tier closed-source models across multiple benchmark tests, sparking what the AI community has called the “DeepSeek shock.”
In late January 2025, DeepSeek’s iPhone app surged past ChatGPT to claim the top spot on the US free app charts, immediately triggering a selloff in US tech stocks. AI chip giant NVIDIA saw its stock plunge 17% in a single day, erasing approximately $600 billion in market value (roughly NT$19.5 trillion, equivalent to the market cap of two Toyotas).
The Japanese government issued an internal notice to various ministries and agencies, requiring full awareness of risks when using generative AI such as DeepSeek. The Personal Information Protection Commission specifically pointed out that personal data obtained by DeepSeek is stored on servers in China. Major Japanese companies such as Toyota Motor, Mitsubishi Heavy Industries, and SoftBank immediately imposed a complete ban on employees using DeepSeek internally. The Taiwanese government also completely banned government agencies from using DeepSeek on national security grounds.
Satirically, at this very moment, one of Japan’s largest-scale models is running this very “untouchable” DeepSeek V3, simply renamed and repackaged as “Japan’s strongest domestic AI” to be presented to the Japanese public.
Their own official website doesn’t even use their own model? The credibility issue is spreading
What’s even more embarrassing is that Japanese netizens have also discovered that Rakuten AI’s official website is likely not actually running their own released Rakuten AI 3.0, but rather Claude or GPT series models. This is because DeepSeek V3 itself doesn’t support image input (it’s not multimodal), yet the Rakuten AI official website does have this feature. Their so-called “strongest AI” that they claimed to release isn’t even being used on their own official website.
As of now, Rakuten Group has yet to make any public statement or response.
Source: KOCPC Chinese