Sakana AI Verifies Fugu's Conductor Model Runs on Gemma 4

Sakana AI published results showing that the “conductor” model at the core of its Fugu multi-agent orchestration system can be trained on Google’s Gemma 4 instead of its original Qwen base, while keeping accuracy and cost efficiency comparable — validating that Fugu’s architecture is not tied to a single base model.

Details

• What Fugu’s conductor does: Fugu operates as a two-layer multi-agent system where a small “conductor” model doesn’t try to hold all knowledge itself, but instead decides how to route each request among a pool of larger, specialized models
• What changed: Sakana AI retrained the conductor using Gemma 4 E2B, Google’s Apache 2.0-licensed small model, in place of the Qwen-based conductor used in earlier Fugu releases
• How it was tested: The company evaluated the new Gemma 4-based conductor against a proprietary test set covering knowledge questions, code correction, code generation, and graduate-level science questions
• Results: The Gemma 4-based conductor achieved accuracy roughly on par with the existing Qwen-based conductor, at comparable cost, while the pool of larger models it routes to remains independently configurable regardless of provider
• Why a small conductor helps: Because the conductor itself can stay small, Sakana AI says it can be retrained and re-validated repeatedly at realistic cost when swapping in a different base model — something that would be far more expensive with a large monolithic model

What happened next

Sakana AI frames the result as evidence that Fugu’s orchestration layer can support “sovereignty” requirements in different markets by swapping in locally trained or open-weight base models — including, potentially, Japanese-developed models — without rebuilding the whole system. The announcement follows Sakana AI’s launch of the Namazu Japanese-language LLM API on August 3 and its Daiwa Securities wealth-management partnership on August 5, continuing a run of releases aimed at demonstrating flexible, provider-agnostic infrastructure for enterprise and sovereign AI customers.