
Sakana AI rebuilds Fugu's conductor model on Gemma 4, matching the Qwen version
Sakana AI retrained Sakana Fugu's conductor — the small model that routes requests to a pool of frontier models — on Gemma 4 E2B, and reports accuracy and cost reduction on par with the existing Qwen-based conductor across held-out knowledge, code, and graduate-level science question sets. The swap shows the orchestration layer is modular too, not the model pool alone, clearing the way for conductors trained on domestically pre-trained Japanese base models so customers with sovereignty requirements can keep reaching top international models through the pool.
Source: sakana.ai ↗
指揮者モデル自身のベースモデルの多様性はこれまで担保できていませんでした。
Sakana AI
Why this matters
- → Proves orchestration layer itself is swappable, not locked to one foundation model.
- → Enables sovereignty-compliant conductor models using domestic Japanese base models.
- → Maintains performance and cost efficiency across different orchestrator generations.
Orchestration goes modular