On October 26, 2026, Sakana AI, Japan’s largest AI unicorn startup, announced plans to develop a completely new domestic AI model to reduce the risks of technological dependence on the US and China.
Provisionally named Sakana Sawara, the model will be entirely self-developed by the company, including the pre-training phase on large volumes of data, rather than relying on foreign models.
According to CEO David Ha, development is set to begin in November 2026, with a small-scale version slated for release in spring 2027, followed by an expansion to a larger model.
The large version is projected to have a scale of hundreds of billions of parameters, though it may still be smaller than the most advanced AI models from the US and China.
Previously, Sakana AI released Sakana Fugu in June 2026, which combined multiple AI models and operated domestically, but still relied partially on foreign model technology.
Sawara aims for deeper autonomy, not only in operational infrastructure but also in model architecture and training processes.
Sakana AI focuses on sensitive fields such as defense, finance, and cybersecurity. In July 2026, the company signed a trial contract with the Japan Ministry of Defense regarding AI-based intelligence analysis.
In cybersecurity, AI is increasingly capable of detecting software vulnerabilities, supporting the remediation of weak points, and analyzing cyberattacks, while also increasing the capabilities of attacking parties.
Developing domestic AI is expected to help Japan better control technology, data, and deployment capabilities within high-security systems.
📌 Conclusion: Sakana AI is shifting from sovereign AI models based on domestic operational infrastructure to full autonomy in both development and pre-training processes. Sawara is scheduled to begin development in November 2026,launch its first version in spring 2027, and scale up to hundreds of billions of parameters. This strategy reflects the trend of nations prioritizing control over core AI technology, especially in defense and cybersecurity, rather than merely hosting foreign models on domestic infrastructure.
