Sakana AI Launches Sovereign Japanese Foundation Model Project Sawara for Defense and Cybersecurity to Reduce US and China Reliance
サカナAIが国産AI開発へ 防衛・サイバーに活用、米中依存リスク排除
Context
Building domestic Japanese foundation models from scratch alongside post training and multi model orchestration marks a strategic pivot by Tokyo's leading AI unicorn to supply sovereign defense and cybersecurity systems independent of American or Chinese providers.
Tokyo based Sakana AI, founded by former Google Brain researchers David Ha and Llion Jones alongside former Japanese diplomat Ren Ito, previously focused on evolutionary model merging, agent orchestration, and post training rather than pretraining large foundation models from scratch.
Deep dive
Japanese AI unicorn Sakana AI announced on Monday that it is launching a full scale domestic foundation model development effort aimed at Japanese defense, cybersecurity, and enterprise sovereignty to eliminate reliance on United States and Chinese models, as reported by Nikkei and confirmed by Sakana AI following its Tokyo symposium.
In an official post summarizing presentations by chief executive David Ha and chief scientist Takuya Akiba, Sakana AI outlined three technical pillars for national AI sovereignty. First, under a pretraining initiative tentatively named Sakana Sawara, a small team will use AI agents to design, prioritize, and execute pretraining experiments for next generation foundation models under constrained compute budgets. Second, its post training program, Sakana Namazu, adapts large language models to Japanese language nuances and specialized business workflows. Third, its orchestration layer, Sakana Fugu, combines multiple models so critical systems never depend on a single vendor.
Sakana AI explained that while it previously focused on the agent and orchestration layers, falling pretraining costs and reinforcement learning specialization now make it vital for Japan to own the underlying model weights for national security and cyber defense.
What to expect next
Japanese government ministries and defense contractors will evaluate Sakana Sawara, Namazu, and Fugu for classified and sovereign workloads, while researchers watch how effectively Sakana's AI agents automate pretraining experiment selection.Industry analysts expect concrete adoption milestones to emerge over the coming quarters as initial deployments scale across production environments. Other allied nations seeking sovereign AI capabilities for defense and critical infrastructure will watch whether Sakana AI's agent driven pretraining approach can deliver competitive domestic foundation models without tens of billions of dollars in compute spending.