Sakana AI will add Nvidia Nemotron models to its Fugu orchestrator
The Tokyo startup says Nvidia’s open-weight models will serve as specialists inside Fugu, though it has not disclosed timing or new benchmark results.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
Sakana AI is adding Nvidia’s open-weight Nemotron models to Fugu, its multi-model AI orchestrator, in an effort to show that coordinated model pools can compete with frontier systems. The companies did not disclose commercial terms, a release date, or new performance figures for the combined system.
Fugu is designed to sit behind a single API while routing work across multiple language models. Sakana AI says the system can break tasks into parts, choose models from an agent pool, assign subtasks, and merge the outputs into one response. The pool can include instances of Fugu itself, and the company has pitched the architecture as a way to reduce reliance on any one model vendor.
The Nemotron models will be used as specialists rather than replacements for the frontier models already available to Fugu. Sakana AI says Nvidia’s models are particularly useful for programming, tool use, and instruction following. That positioning is consistent with the broader agent market, where open-weight models are often judged less by general chatbot performance and more by how reliably they perform narrow work inside larger systems.
Nemotron gives Fugu a larger specialist pool
Nvidia has expanded Nemotron into a family of open models and related tools. The lineup includes Nemotron 3 Ultra, described as having roughly 550 billion parameters and 55 billion active parameters. Benchmark platform Artificial Analysis has called it the strongest open U.S. model to date, ranking it ahead of Google’s Gemma 4 31B, OpenAI’s gpt-oss-120b and Nvidia’s Nemotron 3 Super, while still placing it behind Chinese open-weight models such as Kimi K2.6.
Nvidia has also released Nemotron 3 Nano Omni, a multimodal model that works with text, images, video and audio. That model is aimed at agentic uses including document processing and computer-use agents, according to Nvidia’s positioning of the product.
For Sakana AI, the practical benefit is access to a wider set of models that can be selected for narrower jobs within Fugu. For Nvidia, the integration gives Nemotron another venue to be tested in multi-agent workflows rather than as standalone models. Sakana AI said Nvidia will provide technical guidance on Nemotron recipes and evaluation, and that both teams plan to continue monitoring and optimizing performance after integration.
No new benchmark numbers disclosed
Sakana AI has previously claimed that Fugu Ultra, a stronger version of its orchestrator, performed on par with Anthropic’s Fable 5 and Mythos Preview in the company’s own testing. Early outside tests were more cautious, with concerns about latency and cost. The new Nemotron announcement does not include fresh benchmark data showing how much Nvidia’s models improve Fugu, if at all.
That gap matters because orchestration systems add complexity. Routing across multiple models can improve coverage and resilience, but it can also increase costs, slow responses and make evaluation harder. Sakana AI’s argument is that open models become more useful when coordinated inside agent systems instead of deployed one by one.
The company also frames the approach as a hedge against dependence on a single AI provider, including exposure to outages, policy changes and access restrictions. Sakana AI was founded in Tokyo in 2023 by former Google researchers Llion Jones, a co-author of the Transformer paper “Attention Is All You Need,” and David Ha. The company has centered its research strategy on collective intelligence and has also formed the RSI Lab, which focuses on recursive self-improvement in AI development.
This story draws on original reporting from The Decoder.
