Sakana AI just unveiled Fugu, a model that doesn’t generate answers itself but directs a fleet of other top‑tier models to solve hard problems.
The Simple Version
Think of Fugu as a smart conductor that picks the best musicians for each piece of music.
It sends your request to the most suitable model, gathers the outputs, and blends them into a final answer.
How It Actually Works
Swarm Routing
Fugu first analyzes the task and selects a subset of models that excel at that type of work.
Each selected model runs in parallel, then Fugu aggregates their results using a weighted voting scheme.
Benchmark Results
On SWE‑Bench Pro, Fugu Ultra scores 78% compared to 71% for GPT‑4 and 69% for Anthropic Opus.
The gap widens on multi‑step reasoning tasks where coordination matters.
Technical Deep Dive
Fugu uses the Trinity and Conductor patterns described in Sakana’s research papers.
These patterns define a three‑stage pipeline: query decomposition, model selection, and result fusion.
The API accepts a single JSON payload and returns a unified response, hiding the underlying swarm.
Latency is kept low by caching model routing decisions and by streaming partial results.
Why This Matters
Most frontier models are locked behind closed APIs, forcing developers to depend on a single vendor.
Fugu’s open‑swarm approach lets researchers mix and match models, reducing cost and increasing flexibility.
It also mitigates the risk of sudden bans or policy changes that can cripple a single provider.
Key Takeaways
- Fugu coordinates multiple AI models like a conductor leading an orchestra.
- It achieved higher SWE‑Bench scores than GPT‑4 and Opus in early tests.
- The system avoids vendor lock‑in by using an open routing API.
- Trinity and Conductor strategies enable efficient model selection and fusion.
- Latency is managed through caching and streaming of partial outputs.
- Open swarm architectures could reshape how enterprises deploy AI.
- The black‑box nature of aggregated results still requires careful validation.