Sakana AI has integrated Nvidia's Nemotron models into its Fugu orchestrator, aiming to demonstrate that collective intelligence can rival single frontier models. The Fugu system dynamically selects suitable models for subtasks and aggregates their results, with Nemotron acting as a specialist in programming, tool usage, and instruction following. According to Sakana AI, this approach reduces dependence on individual vendors and enhances flexibility by allowing new models to be added at any time. The company emphasized that the system is modular and not tied to the strengths or outages of any single provider. Source: thedecoder
Nvidia's Nemotron family includes open-weight models and tools, with Nemotron 3 Ultra being highlighted as the most capable open US model to date. This model has roughly 550 billion parameters and 55 billion active parameters, outperforming Gemma 4 31B and gpt-oss-120b, though it still trails behind Chinese models like Kimi K2.6. Additionally, the Nemotron 3 Nano Omni, a multimodal model, handles text, images, video, and audio, targeting agentic use cases such as document processing and computer-use agents. Together, these models provide a broad range of capabilities that Fugu can draw from when selecting agents. Source: thedecoder
Sakana AI's Fugu system is designed to behave like a single model externally while internally orchestrating multiple language models. The company claims its stronger variant, Fugu Ultra, performed on par with Anthropic's Fable 5 and Mythos Preview, though early independent tests were less enthusiastic, citing issues with speed and cost. The integration of Nemotron is expected in an upcoming Fugu release, with ongoing monitoring and optimization planned. Nvidia will provide technical guidance on Nemotron recipes and evaluation. Source: thedecoder
