Sakana AI Names Sawara as Project for Domestic Model Pretraining

Sakana AI is moving to pretrain a model from scratch through its Sakana Sawara project, adding in-house development to a strategy that previously emphasized combining existing open-source models.

The company named Sawara as the provisional pretraining project on Oct. 26. It places the project alongside post-training through Sakana Namazu and orchestration through Sakana Fugu in a three-part AI strategy.

Sakana cited lower model-development costs, a narrowing gap between open models and leading systems, and reinforcement learning's potential to tailor models to specific tasks as reasons to pursue pretraining. The company also identified computing resources as a constraint on the work.

Sakana's earlier approach focused on combining existing open-source models through evolutionary techniques. The new pretraining effort expands that strategy to include developing a model from scratch.

The company described using AI agents to help plan pretraining experiments. The project is part of the pretraining pillar, with Namazu and Fugu covering the other two parts of Sakana's strategy.