Sakana Evolutionary Model Merge Advantage

The real edge in Sakana’s approach is speed to a usable model, not bragging rights on parameter count. Incumbents like SoftBank and Preferred are building expensive base models from scratch, which means long GPU booking, long data pipelines, and long tuning cycles before a customer can use anything. Sakana starts with models that already exist, then searches for better combinations, so it can ship a Japan specific model or workflow much faster and with far less capital tied up.

• SoftBank’s Sarashina effort is built around very large training runs. SoftBank has described plans ranging from 390 billion to 1 trillion parameters and has built a Blackwell based DGX SuperPOD to support them. That kind of program needs major capex before any product revenue shows up.
  2 softbank 3 softbank
• Preferred’s PLaMo follows the same basic playbook, training a domestic model from scratch and then packaging variants for finance, translation, cloud API, and on premises use. That works well for regulated buyers, but every new domain still depends on a heavyweight base model program underneath.
  4 preferred 5 preferred 7 preferred
• Sakana’s Evolutionary Model Merge changes the bottleneck from pretraining to search. Users pick a pool of existing models and a benchmark, then the system generates and tests many child models over multiple rounds. Sakana says this process produced a 7B Japanese model that beat some earlier 70B Japanese models on benchmarks.
  1 sacra 6 sakana

This points toward a split market. The largest incumbents will keep owning giant national models and secure infrastructure, while Sakana is positioned to win where customers care more about fast specialization, lower compute bills, and getting a good model this quarter instead of training a perfect one next year.