Sakana AI pairs Nemotron with Fugu orchestrator to show open models can match frontier AI

Sakana AI, founded in Tokyo in 2023 by former Google researchers Llion Jones (co-author of "Attention Is All You Need") and David Ha, launched Fugu as a system that orchestrates multiple language models through a single API. Fugu itself is a language model trained to call other LLMs from an agent pool that includes instances of itself, dynamically selecting and combining models for given tasks, delegating subtasks, and synthesizing results. The architecture is deliberately modular: new models can be added at any time without tying the system to the strengths or outages of any single provider.

In its own benchmarks, Sakana AI claimed that Fugu Ultra—the stronger variant—performed on par with Anthropic's Claude 3.5 Sonnet and Mythos Preview. However, early independent tests were less enthusiastic, with criticism around speed and cost. The new partnership with Nvidia aims to strengthen this orchestration strategy by adding Nemotron models to the agent pool. Nvidia's Nemotron family consists of open-weight models and tools; Sakana AI specifically points to their strengths in coding, tool calling, and instruction following. These specialist models are designed to complement frontier models inside Fugu's orchestration layer, not replace them. Nvidia's Nemotron 3 Ultra features roughly 550 billion parameters and 55 billion active parameters; according to benchmark platform Artificial Analysis, it is the most capable open US model to date, ranking ahead of Gemma 4 31B and Nvidia's own Nemotron 3 Super, though still trailing Chinese models like Kimi K2.6. The company also recently shipped Nemotron 3 Nano Omni, a multimodal model handling text, images, video, and audio, aimed at agentic use cases like document processing and computer-use agents.

Sakana AI has not announced a specific date for the integration, only that it will ship in an upcoming Fugu release. After launch, the Sakana and Nemotron teams plan to monitor and optimize Nemotron's performance inside Fugu on an ongoing basis, with Nvidia providing technical guidance on Nemotron recipes and evaluation. The partnership carries no new benchmark results yet. Sakana AI frames the move as part of a broader trend: progress in AI will increasingly depend on how well models can be evaluated, combined, and woven into real-world workflows. No single model will lead in every task, language, modality, and enterprise environment, making the orchestration layer a critical piece of the next phase of open AI. The company positions this as a Japanese "collective intelligence" approach designed to give developers and companies worldwide access to a growing ecosystem of open models. From its founding, Sakana AI has centered collective intelligence and orchestration—rather than ever-larger single models—at the core of its scaling strategy. Before Fugu, the company had set up the RSI Lab, a research group focused on recursive self-improvement aimed at automating the AI development process itself.

Sakana AI's decision to integrate Nemotron reflects a strategic shift in how open-source AI systems are being positioned. Rather than competing head-to-head with frontier closed models like OpenAI's GPT or Anthropic's Claude, the startup is betting that orchestration—dynamically routing tasks to specialized models and synthesizing results—can deliver competitive performance while preserving modularity and reducing vendor lock-in. This approach addresses real constraints: independent evaluation of Fugu has highlighted speed and cost concerns, which suggests that a single monolithic open model may not yet match frontier systems in both quality and efficiency. By adding Nvidia's Nemotron family, which covers coding, tool calling, multimodal reasoning, and parameter-efficient inference, Sakana AI broadens the range of specialist agents available to Fugu's orchestrator.

The partnership also reflects Nvidia's strategy to expand Nemotron as a competitive open alternative to closed models, complete with tooling and recipes. Nvidia's willingness to provide technical guidance and collect performance data from Fugu's production use cases suggests both companies see value in understanding how open models perform in multi-agent workflows—a signal that agentic systems and orchestration are becoming a central axis of AI development. Sakana AI frames this move in geopolitical and risk-management terms: open, orchestrable models reduce dependence on any single provider and offer a hedge against regulatory or foreign-policy barriers to access. This framing is particularly resonant for international companies and developers seeking alternatives to closed US-headquartered AI providers.

No comments yet. Be the first to share your thoughts!