Did *Tekken* really make everyone laugh? Is Rakuten AI 3.0 being exposed as just a rebranded product?

Something feels off.

The core of the controversy around Rakuten AI 3.0 this time can be summed up in one sentence: the promotion made it look like an in-house model, but the community quickly dug up that the foundation was actually a fine-tuned version of DeepSeek V3. On the 17th, Japan’s Rakuten Group announced Rakuten AI 3.0 with great fanfare, even branding it as “the largest in Japan.” But in less than 12 hours, the tech community followed the public documents and saw right through it.

So where was the problem? The evidence was just too obvious.

People checked the model files Rakuten uploaded to Hugging Face that day and found that in config.json, "architectures" was listed as "DeepseekV3ForCausalLM" and "model_type" as "deepseek_v3". This was not even skating close to the line—it was like they did not even bother cleaning up the startup animation. Looking at the parameters, Rakuten AI 3.0 had 671B total parameters and 37B active parameters, exactly matching the original DeepSeek V3. By the original description, this was essentially a re-skinned fine-tune, with no core modifications.

Let’s be clear first. On the legal side, this may not necessarily violate any rules. DeepSeek V3 uses the Apache 2.0 license, which allows commercial use. In other words, taking it for secondary development is itself perfectly allowed. The real reason people were angry was not that it “used open source,” but that Rakuten only said it had “fully leveraged the best achievements of the open-source community” in its promotion, without explicitly naming DeepSeek, while the overall packaging strongly suggested independent development.

It is like using someone else’s pre-made combo in a match, changing only the finisher, and then going on stage claiming it is your own original system. By the rules, that might be fine. In terms of public perception, it is guaranteed to blow up.

Now for the part that hurt the optics even more. Some users found that Rakuten had initially removed DeepSeek’s license files, and only re-uploaded them later as a “NOTICE” file after being called out. That move was fatal. If they had been upfront from the start, the community might not have gone after them so hard. But if you hide it first and patch it later, the whole thing leaves a different taste.

Japanese netizens were especially angry this time, and it is not hard to see why. On one hand, Rakuten was accused of taking subsidies from the Japanese government, only to present a fine-tuned version of a Chinese open-source model in the end. On the other hand, this is not an isolated case. Nikkei News had previously reported that among the top 10 AI models developed by Japanese companies, 6 were based on DeepSeek or Qwen for secondary development. That shows what the current industry ecosystem looks like: base models are becoming increasingly concentrated, while differentiation mainly comes from fine-tuning, deployment, and scenario-specific adaptation.

My view is very straightforward. Using open source is nothing to be ashamed of. The open-source community exists precisely so people can chain combos together and optimize them. What is embarrassing is benefiting from open source while dressing up the messaging as if it were a fully self-developed first release, like trying to steal frames.

This incident is also a reminder to the industry. The next time a major company releases an AI model, do not start by listening to the poster copy—look first at the model card, license, config, and parameter sheet. Those are the hard data. Marketing copy can perform. Fields do not.

One final point. For ordinary users and gamer communities, this kind of controversy does not just affect a single launch event—it damages long-term trust. The AI space right now is as intense as a high-stakes match: performance is only the first round, while transparency is the endgame that follows. Whoever clearly explains the model’s origins is the one who can truly stand firm.