Fugu Ultra sits inside Sakana AI's Fugu family of orchestrator language models, a design where the model itself is trained to read a user query and then dynamically construct an agentic scaffold that coordinates a team of other LLMs rather than answering alone. According to the Sakana Fugu Technical Report authored by the Fugu Team at Sakana AI, this orchestrator framing is meant to amplify the specializations of many frontier models into one collectively intelligent system. The Ultra variant is positioned for the hardest problems, trading latency for answer quality, while a sibling Fugu model targets more routine, latency-sensitive use.
The training paradigm described in the report combines large-scale fine-tuning, evolutionary algorithms, and reinforcement learning, which together teach the orchestrator when to invoke which sub-agent and how to stitch their outputs into a coherent response. Sakana reports that this approach reached state-of-the-art results across demanding public benchmarks including SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. In practice, Fugu Ultra is best suited to complex, multi-step reasoning and coding workflows where composing several specialized models yields better outcomes than any single LLM, with the scaffold-making step providing a natural fit for tool use and structured intermediate results.
