Sakana AI Adds NVIDIA's Nemotron 3 Ultra to Fugu, Proving Open Models Can Route Past Proprietary Walls

Sakana AI is expanding its open-source model orchestration platform Fugu with the addition of NVIDIA's Nemotron 3 Ultra, widely considered one of the most capable open-weight models from a US lab. As reported by Tech Times, the integration deepens the case that open models, when routed intelligently, can compete with and even surpass proprietary alternatives.

Fugu is a novel AI system developed by Sakana AI that uses a smart routing algorithm to dynamically select the best combination of open-source models for each task, rather than relying on a single monolithic model. The approach takes inspiration from the Japanese pufferfish dish it is named after — combining the best parts of different ingredients into a superior whole.

Nemotron 3 Ultra is NVIDIA's flagship open-source large language model. By adding it to the Fugu routing pool, the system can now intelligently decide which queries are best handled by Nemotron and which should be directed to other open models in the pool, creating an optimized outcome for each request.

The partnership addresses a fundamental question in the AI industry: can open-source models truly replace proprietary systems? Fugu's orchestration approach suggests the answer is yes — not by matching a single model's peak capability, but by leveraging ensemble intelligence to achieve superior aggregate performance.

Sakana AI, founded by former Google researchers and headquartered in Tokyo, has been gaining significant traction in the open-source community with its Fugu project. The NVIDIA collaboration adds both technical credibility and access to one of the most powerful open-weight model families available.

From an industry perspective, the development signals a shift in how organizations think about AI infrastructure. Instead of committing to a single model vendor, enterprises can maintain a diverse model pool and route tasks based on performance characteristics and cost constraints. This model-routing architecture is increasingly seen as a pragmatic path forward for AI deployment.

The broader implication is that the open-source versus proprietary debate is evolving from a binary choice into a systems-design question: the right combination of open models, orchestrated intelligently, can deliver outcomes that challenge the dominance of any single proprietary system.