Sakana AI will build domestic AI models from scratch in Japan

Sakana AI made its name by taking other people’s AI models and making them better at being Japanese. Now the Tokyo startup will build domestic AI models from scratch. From remixing to recording originals Until now, Sakana AI’s playbook has rested on two techniques. The first is post-training, which means taking an existing foundation model and refining it for specific tasks and audiences. The second is something the company calls evolutionary model merging. Think of it as selective breeding for software: combine traits from different models, keep what works, and discard the rest. The approach has been sold on efficiency and cultural alignment. Rather than spending heavily to train a giant model from zero, Sakana adapted what already existed and tuned it for Japanese language, values, and use cases. That strategy set Sakana apart from peers who chose to build AI frameworks from the ground up. The new plan puts the company closer to that camp, at least for its domestic models. Sakana has pitched itself as a hedge against reliance on foreign models, particularly in sensitive sectors like finance and defense. A fast-growing company with deep backing Sakana AI was founded in Tokyo in 20...