Sakana must move up the stack

This pressure means Sakana cannot win by making model merging merely possible, it has to win by making merged models measurably better for real enterprise jobs. Open tools like mergekit already let developers combine pretrained models, and Sakana itself has open sourced its evolutionary model merge work, so the durable value shifts from the merge mechanic to the full workflow, search, evaluation, and deployment layer wrapped around it.

• mergekit turns model merging into a developer utility. It provides open tools for merging pretrained LLMs, which lowers the barrier for anyone to reproduce a basic version of Sakana's core technique without paying for a proprietary platform.
  1 github
• Sakana's own research shows the harder part is not averaging weights, but running many generations, scoring candidates, and selecting the best recipes for a task. That creates room to sell a system that automates search and benchmarking for banks or other enterprises.
  2 sakana 4 sacra
• The MUFG partnership shows where pricing power can still live. A bank is not buying a merge algorithm by itself, it is buying a working document automation system, tuned models, implementation support, and integration into a regulated workflow.
  3 sakana 4 sacra

Going forward, model merging will likely look like an open ingredient, not a standalone product. Sakana's path is to move up the stack into domain specific systems, managed optimization, and enterprise orchestration, where customers pay for better outcomes in credit, research, and operations rather than for the merge itself.