Japan plans to use Fugu as "domestic AI" in the Self-Defense Force, but the command model has always been Qwen.

According to Dynamic Insight monitoring, the Japanese government is considering using Sakana Fugu for Self-Defense Force command and control. The Yomiuri Shimbun even lists it as one of the main candidates for "domestic AI." The defense application also requires the exclusion of Chinese AI, and related developers must meet Japan's security screening requirements.

However, this "domestic AI" is quite subtle. Fugu itself is more like a set of AI scheduling systems: a small-scale command model is responsible for task allocation, while the actual work is done by larger closed-source models like GPT, Claude, and Gemini. Today, Sakana explicitly admitted that in the past, this command model has always been based on Alibaba's Qwen training.

If Fugu is to be used for defense purposes, this layer of Qwen will have to be replaced. Sakana has already retrained the command model with Google DeepMind's open-source Gemma 4 E2B, with almost no drop in performance. The next step is to switch to a domestic model that has been entirely developed in Japan since the pre-training stage.

Japan is working hard to localize the "command center" of Fugu, but the models that truly provide cutting-edge capabilities still come from the United States. This system, which is regarded as a representative of "domestic AI," may ultimately end up with Japanese models in charge of command, American models doing the work, and Chinese models being kept out.