Sakana AI NVIDIA Partnership: Nemotron Joins Fugu

Sakana AI and NVIDIA bring Nemotron into Sakana Fugu

Nemotron, NVIDIA's open model family, will join Sakana Fugu as specialized agents in an upcoming version of Fugu. Sakana AI announced this in a post titled "Sakana AI Teams With NVIDIA to Advance Open Model Innovation from Japan." The post describes Nemotron as a model family that comes with open weights and tooling.

Sakana AI frames the work as a joint effort. Its co-founder and CEO, David Ha, says the company wants to build the next generation of Fugu orchestration models together with NVIDIA, by incorporating leading open-weights models like Nemotron. The post states no investment amount.

How Fugu handles AI orchestration behind one API

AI orchestration means one system picks and combines several AI models for a job, instead of sending everything to a single model. Sakana Fugu works this way: a developer sees one interface and one API, and behind it Fugu selects, coordinates and merges the models and agents that suit each task into one response.

The design is modular. New and better models can be added over time, so Fugu improves with the wider AI ecosystem. It is also never tied to the strengths, limits or availability of any single model, which the post says makes Fugu more adaptable and resilient.

Fugu is itself a language model. It was trained to call other LLMs from an agent pool, including copies of itself, so it can work through complex tasks that take many steps. For the product on its own, see Sakana Fugu Ultra is not always the best Fugu.

What NVIDIA Nemotron adds as a specialized agent

Nemotron models will add new capabilities to Fugu's agent pool, next to the other open and specialized models Fugu orchestrates. The post says Nemotron can complement frontier models inside the orchestration. Kari Briski, NVIDIA's Vice President of Generative AI, says open models let countries build AI that reflects their own language, culture and policies. She adds that Nemotron in Fugu shows what open and closed models can do when they are orchestrated well.

Open model strengths in coding and tool calling

The post names three strengths that set Nemotron apart: coding, tool calling and instruction following. Sakana AI uses Nemotron to make a wider point: open models become more useful when they are orchestrated together inside agentic systems than when each one is used alone.

How NVIDIA supports Sakana Fugu with recipes and evaluation

The work starts after Nemotron is integrated as a specialized agent in an upcoming version of Fugu. From then on, the Sakana and NVIDIA Nemotron teams will keep observing how Nemotron performs in Fugu and work to improve it.

NVIDIA's part is technical guidance on Nemotron recipes and evaluation best practices. The aim is to help Fugu get the strongest performance out of Nemotron in its multi-agent workflows. Both companies will then evaluate Nemotron inside those workflows and use the findings to improve both Nemotron and Fugu.

Why Sakana AI calls orchestration the next critical layer

Sakana AI argues from a trend in open models. As they become more numerous, more capable and more specialized, progress depends on how well they can be evaluated, combined and deployed in real work. No single model is likely to hold every advantage across every task, language, modality and company environment. In Sakana AI's argument, orchestration is the layer that turns a varied model ecosystem into practical, reliable systems.

Sakana AI also reports that early evaluations show Fugu performing strongly alongside leading frontier systems. It reads this as a sign that coordinating models can become a way to scale AI in its own right. The next step it names is to widen the diversity, specialization and quality of the models Fugu can use.

The post adds that NVIDIA's open model stack, its support for companies building on open models and its accelerated computing platform can help scale this layer. They would do so by providing high-quality models to orchestrate and the infrastructure to run them efficiently.

The feedback loop between open model use and Fugu

The post calls the collaboration a reinforcing cycle between open models, orchestration and real-world use. It has four links:

• Fugu gets a deeper pool of specialized model capabilities.
• NVIDIA can evaluate how its models perform when they are coordinated in agentic, multi-step workflows.
• Developers and enterprises get more choice, more flexibility and less dependence on any single model or provider.
• Over time, real-world signals can help improve both the models and the orchestration layer, so everyday usage becomes feedback.

The stated goal is to advance open model innovation from Japan for the world. The post presents this as a Japan-born approach to collective intelligence, that is, many models working in concert.

Did NVIDIA announce any funding or GPU supply?

No. The post states no investment amount, no GPU quantity and no release date for the upcoming Fugu version, and it names no specific Nemotron model. What it describes is technical collaboration: NVIDIA will give guidance on Nemotron recipes and evaluation best practices, and both teams will evaluate Nemotron inside Fugu. It comes down to three points:

• Nemotron will be integrated as a specialized agent in an upcoming version of Fugu.
• NVIDIA will provide technical guidance on Nemotron recipes and evaluation best practices.
• Sakana AI and NVIDIA will evaluate Nemotron in Fugu's multi-agent workflows and feed the results back into both products.

For anyone following the Sakana AI NVIDIA story, the open items are the Fugu version that carries Nemotron and which Nemotron models it uses. The post answers neither.