Rakuten’s “Japan‑largest” AI sparks outrage after netizens say it’s just DeepSeek V3

Overview

Rakuten (楽天グループ/乐天集团) on March 17 unveiled Rakuten AI 3.0, billed as “Japan’s largest high‑performance AI model” with roughly 700 billion parameters and supported by subsidies from Japan’s METI and NEDO under the GENIAC programme. It has been reported that Japanese users quickly flagged that the model’s config on Hugging Face identifies it as DeepSeek V3, an open‑source Chinese model, and that Rakuten omitted explicit credit to DeepSeek in its launch materials.

What netizens found

Users inspecting the uploaded weights reportedly found config.json entries like "model_type":"deepseek_v3" and architecture names that match DeepSeek V3 exactly; the published layer and expert counts line up with DeepSeek’s 671 billion‑parameter configuration. Critics also pointed out Rakuten initially did not include DeepSeek’s MIT licence file in the repository and only added a NOTICE with a copyright line after being called out. Legally, using and fine‑tuning an open‑source model is permitted, but netizens say the problem is transparency — and tone.

Why this matters

This is sensitive politics as much as it is engineering. GENIAC is a government‑backed effort to foster Japan’s autonomous AI capability by subsidising compute; taxpayers’ money was involved. Many Japanese observers see a reputational problem: subsidised “national” AI built on a model that many companies and governments have treated as high‑risk or banned outright. It has been reported that DeepSeek earlier triggered international restrictions and concern over data residency and national security, so the provenance of core weights matters far beyond PR.

Broader context

The episode raises a familiar choice for governments and firms: build from scratch at high cost, or import and adapt foreign open‑source models. Historically, Japan’s response to disruptive foreign tech has oscillated between rejection and “adopt the technique to beat the technique.” Rakuten’s handling — ambiguous crediting, late licence patching, and a national‑pride framing — has made that trade‑off a public controversy rather than a quiet technical decision.