[Japan Plans to Use Fugu as 'Domestic AI' for Self-Defense Forces, But the Command Model Has Always Been Qwen] According to monitoring by Beating, the Japanese government is considering using Sakana Fugu for command and control within the Self-Defense Forces. *Yomiuri Shimbun* has even listed it as one of the main candidates for 'domestic AI.' For defense purposes, the system must exclude Chinese AI, and relevant developers must meet Japan's security review requirements. However, this 'domestic AI' is actually quite nuanced. Fugu itself is more like an AI dispatch system: a small command model is responsible for task allocation, while the actual work is carried out by backend closed-source large models such as GPT, Claude, and Gemini. Sakana explicitly acknowledged today that this command model has historically been trained based on Alibaba's Qwen. If Fugu is truly to be used for defense purposes, the Qwen layer will need to be replaced. Sakana has already retrained the command model using Google's DeepMind open-source Gemma 4 E2B, with performance remaining largely intact. The next step involves switching to a domestically developed model trained entirely in Japan from the pre-training stage. Japan is striving to localize the 'command tower' of Fugu, but the models providing cutting-edge capabilities are still primarily from the United States. This system, touted as a representative of 'domestic AI,' may ultimately end up as follows: Japanese models handle command, American models do the heavy lifting, and Chinese models are barred from entry. [Original Link]