Japan's Fugu Defense AI Replaces Qwen Command Layer With Open-Source Gemma Model

Sakana retrained Fugu's command model using Google's Gemma to remove Alibaba's Qwen, aiming for domestic pre-training while retaining US large models for core tasks.

Woofun AI reports that the Japanese government is evaluating Sakana Fugu for Self-Defense Forces command and control, categorizing it as "domestically developed AI." The system functions as an orchestration layer, delegating actual processing to closed-source models like GPT, Claude, and Gemini. Sakana disclosed that the original command model was trained on Alibaba's Qwen, necessitating a replacement to satisfy security reviews excluding Chinese technology. The team has retrained this component using Google DeepMind's open-source Gemma 4 E2B, maintaining performance levels. Future plans involve transitioning to a model pre-trained within Japan, resulting in a hybrid architecture where Japanese-led command coordinates US-based execution capabilities.