Tokyo-based startup Sakana AI is adding Nvidia’s open Nemotron models to its Fugu orchestrator. The partnership aims to demonstrate that coordinated open models can compete with frontier systems.
Sakana AI launched Fugu recently. The system is a language model trained to call other LLMs from an agent pool that includes instances of itself. Behind a single API, Fugu dynamically picks which models to combine for a given task, delegates subtasks, and synthesises the results into one response.
The setup is modular. New models can be added at any time, so the system is not tied to the strengths or outages of any single provider, according to Sakana AI. In its own benchmarks, the company claimed its stronger variant Fugu Ultra performed on par with Anthropic‘s Fable 5 and Mythos Preview. Early independent tests were less enthusiastic, though, with criticism around speed and cost.
Nemotron fills a specialist role in Fugu’s agent pool
Nvidia’s Nemotron family consists of open-weight models and tools. Sakana AI points to their strengths in coding, tool calling, and instruction following. As specialist models, they are meant to complement the frontier models inside Fugu’s orchestration layer, not replace them. Open models become more useful when they are orchestrated in agentic systems rather than deployed in isolation, the company says.
Nvidia has been expanding the Nemotron lineup fast. With Nemotron 3 Ultra, a model with roughly 550 billion parameters and 55 billion active parameters, the company released what benchmark platform Artificial Analysis calls the most capable open US model to date. It ranks ahead of Gemma 4 31B, gpt-oss-120b, and Nvidia’s own Nemotron 3 Super, but still trails Chinese models like Kimi K2.6.
Nvidia also shipped Nemotron 3 Nano Omni, a multimodal model that handles text, images, video, and audio, aimed at agentic use cases like document processing and computer-use agents. Together, the Nemotron family covers a broad range of capabilities that Fugu can draw from when picking agents.
Sakana AI has not given a specific date for the integration, saying only that it will ship in an upcoming Fugu release. After that, the Sakana and Nemotron teams plan to monitor and optimise Nemotron’s performance inside Fugu on an ongoing basis. Nvidia will provide technical guidance on Nemotron recipes and evaluation.
Orchestration as a scaling path for open AI
Sakana AI frames the partnership as part of a bigger trend. Progress in AI will increasingly depend on how well models can be evaluated, combined, and woven into real-world workflows, the company argues. No single model will lead in every task, language, modality, and enterprise environment. That makes the orchestration layer a critical piece of the next phase of open AI.
“The most capable AI won’t come from any single model, but from many models working in concert,” Sakana AI writes in its announcement.
In early evaluations, the orchestration-based approach showed strong performance alongside leading frontier systems. The announcement does not include any new benchmark numbers for the Nemotron combination, though. In practice, the deal means Sakana gets access to a wider pool of specialist models while Nvidia collects data on how Nemotron performs in multi-agent workflows.
Sakana AI also positions the partnership in geopolitical terms. The startup describes its contribution as a Japanese “collective intelligence” approach designed to give developers and companies worldwide access to a growing ecosystem of open models. When it first unveiled Fugu, the company had already pointed to the risks of depending on a single AI provider and pitched open, orchestrable models as a hedge against regulatory or foreign-policy access restrictions.
The startup was founded in Tokyo in 2023 by former Google researchers Llion Jones, co-author of the Transformer paper “Attention Is All You Need,” and David Ha. From the start, they put collective intelligence rather than ever-larger single models at the center of their scaling strategy. Before Fugu, Sakana AI had set up the RSI Lab, a research group focused on recursive self-improvement that aims to automate the AI development process itself.
What it means
For developers using Fugu, the addition of Nemotron adds specific capabilities to the pool of available agents. Users can now tap into models specialising in coding, tool usage, and instruction following without relying on a single vendor for these functions. This modular approach allows teams to swap in or out models based on specific needs, reducing reliance on any one provider’s uptime or pricing. It also means the system can handle more complex workflows by combining different strengths, though practical speed and cost remain areas to watch.
