Japan’s ‘Self‑Developed’ 700B AI Model: A DeepSeek Re‑skin Flop

On March 17, 2026, Rakuten announced Rakuten AI 3.0 , touting it as Japan’s biggest and most powerful self‑developed large model with roughly 700 billion parameters and claiming support from the Ministry of Economy, Trade and Industry.

Within hours, observant developers opened the model’s config.json on Hugging Face and saw the line: "architectures": ["DeepSeekV3ForCausalLM"] which indicates the model is built on DeepSeek V3, not a wholly original architecture.

Embarrassing Scene

Rakuten’s presentation proclaimed a home‑grown AI, yet the underlying code mirrors DeepSeek. The contrast is illustrated with a humorous analogy: buying flour, making a bun, and then claiming it as a self‑produced staple.

Key Points of the Flop

“Self‑developed” is just a Japanese fine‑tuning of DeepSeek V3. The model’s core remains the open‑source architecture.

Government involvement. The project received funding from Japan’s GENIAC program, highlighting official backing for the publicity.

Broken promises. The launch deck boasted “Japan’s strongest” and “autonomous development,” but the code reveals no original contribution.

Commercial Rationale Behind the Choice

From a business perspective, Rakuten’s decision has some logic:

DeepSeek V3 offers high performance at low cost, comparable to GPT‑4.

Training a 700 billion‑parameter model from scratch would be prohibitively expensive.

Leveraging an existing giant model lets Rakuten “stand on the shoulders of a giant.”

Nevertheless, the company chose to market the fine‑tuned version as a wholly original breakthrough, inviting widespread ridicule.

Psychological Angle: Japanese Pride

The episode reflects a mindset of “I may be weak, but you cannot say I am.” Japan feels pressure as the US and China dominate AI development, and the desire to appear competitive leads to overstated claims.

Faced with two unattractive options—investing massive funds to build from scratch or admitting lag—Rakuten opted for a glossy façade: “We didn’t copy, we deeply optimized open‑source architecture,” and “We integrated global advanced technology,” a classic case of “the art of words.”

Takeaways

Even the most polished PPT cannot hide dishonest code.

True self‑development is hard; exaggeration should be tempered with humility.

Before bragging, verify the model’s configuration files.

Finally, a reminder to Rakuten: close any backdoors before bragging, and avoid tagging the official account when sharing this story.