Sakana AI taps Nvidia Nemotron to prove many models beat one

Sakana AI is betting that the smartest AI won’t be a single model — it’ll be a swarm of them. The Tokyo-based startup is adding Nvidia’s open-source Nemotron model family to Fugu, its recently launched AI orchestrator, in a deal that’s less about raw horsepower and more about proving a philosophy.

Fugu itself is a language model trained to act as a conductor. Behind one API, it dynamically picks which models from a growing agent pool are best suited for subtasks, then synthesizes their outputs into a single response. Sakana claims its top variant, Fugu Ultra, already performs on par with Anthropic’s higher-end systems in its own benchmarks — though early independent testers grumbled about speed and cost.

The Nemotron models will slot in as specialists, not replacements for the heavy-lifting frontier models. Nvidia has been shipping these open-weight models fast. The 550-billion-parameter Nemotron 3 Ultra recently ranked as the most capable open US model on Artificial Analysis, beating out Gemma 4 31B and Nvidia’s own Nemotron 3 Super. It still trails Chinese models like Kimi K2.6, but for coding, tool calling, and following instructions, Sakana sees a sharp set of tools to plug into Fugu’s modular system.

Sakana frames this as more than a technical integration. Founded in 2023 by former Google researchers Llion Jones (a co-author of “Attention Is All You Need”) and David Ha, the company has pushed “collective intelligence” from day one. “The most capable AI won’t come from any single model, but from many models working in concert,” the announcement reads. There’s a geopolitical angle here too — Sakana pitches open, orchestratable models as a hedge against depending on a single vendor or getting caught in regulatory crossfire. No specific launch date or new benchmark numbers came with the news, just a promise that Nemotron will ship in an upcoming Fugu release and that Nvidia will keep tinkering with it afterward.