Key facts
- Sakana AI will develop its own domestic foundation models from scratch.
- The new project is tentatively named Sakana Sawara.
- Development is expected to begin in November 2026.
- A smaller model is slated for release in spring 2027.
- The company aims to reduce reliance on U.S. and Chinese AI models.
- Sawara is intended for sensitive fields like defense and cybersecurity.
Japanese artificial intelligence startup Sakana AI announced on Monday that it will develop its own domestic foundation models from scratch, a move aimed at reducing the risks associated with relying on U.S. and Chinese models, particularly in sensitive sectors like defense and cybersecurity. The company's new project is tentatively named Sakana Sawara. Development is slated to begin in November 2026, with a smaller model expected in spring 2027, followed by larger versions. Sakana AI aims to gain control over the base model's data mixture, tokenizer, and training objectives by starting from randomly initialized weights, offering greater data residency and operational control for Japanese institutions.