Sakana AI

Sakana AI

Merge open-source models into new ones with far less compute, and let an automated pipeline carry a research idea from experiment to paper draft.

Visit website

What is Sakana AI?

A decent set of open-source models on hand, and the goal is one that handles Japanese well — the usual next step is another training run and the compute bill that comes with it. Sakana AI takes a different route: let the models evolve. Its evolutionary model merge treats existing open-source models as building blocks, tries combinations in a process modeled on natural selection, and keeps the variants that perform better, ending up with a new model that inherits strengths from several parents instead of pretraining from scratch.

That approach produced EvoLLM-JP for Japanese text and EvoVLM-JP, which reads both text and images; the weights are published, so they can be downloaded for inference or further fine-tuning. Teams watching their compute budget while wanting a solid Japanese starting point have a reason to look here first. When comparing options on FindAIGo, weighing saved training cost against Japanese-language performance tells you more than parameter counts alone.

Research itself is the other target. The AI Scientist is an automated pipeline: hand it a direction and it generates ideas, writes experiment code, runs the experiments, reads the results, then assembles the work into a paper draft and puts it through an automated review pass. The code is open source, but it expects model API access and compute of your own, and a run takes real time and money to watch over — less a one-click result machine than an assistant that keeps working while you sleep.

Who it suits is fairly clear: researchers working on models, teams that want a particular language capability without a large training bill, and anyone curious how far automated research has come. Before starting, decide whether the goal is a model to use or a framework to modify — inside Sakana AI those are two different doors.

加载中 / Loading

Discussion

Loading comments…