Nikkei reports that the Tokyo company plans to release a smaller model, provisionally called Sakana Sawara, in spring 2027, then develop a larger model.
The company intends to handle pre-training itself rather than only combine external models. The later, larger model is expected to have hundreds of billions of parameters or more. Sakana AI cited potential use in defence, cyber defence, and finance. Its existing Sakana Fugu system, generally available since June, combines multiple models and is operated in Japan.
David Ha, Sakana AI's CEO, described the effort as training a model from scratch. The report says the company wants to use AI in its development process to improve efficiency and reduce costs.
The announcement does not establish the model architecture, training data, compute supplier, benchmark performance, exact parameter count, release terms, price, API access, or whether the stated spring 2027 schedule will be met. Source: https://www.nikkei.com/article/DGXZQOUC261H80W6A021C2000000