ベースモデルに依存しないオーケストレーションに向けて：Gemma 4版 Sakana Fuguの検証

Sakana AI reports that Fugu's coordinating model can be rebuilt on Google DeepMind's Gemma 4 E2B. Fugu separates a small model that decides how to assign work from a pool of models that performs it. In this experiment, the new coordinator used the same training approach as the existing Qwen-based version.

On Sakana's held-out internal question set, the company reports that the Gemma-based coordinator matched the existing coordinator's accuracy and delivered a comparable cost reduction relative to random routing among the pool's models. That supports treating the coordinator's foundation as another configuration choice. Sakana presents a future coordinator built on a domestically pretrained model as a development goal; the experiment itself does not establish a wholly domestic processing chain.