🚀 Sakana Fugu: A Family of Orchestrator Models

Photo by Jo Lin on Unsplash

If you have been following the progress of large language models over the last few years, you have probably noticed something interesting. Every few months, a new model arrives claiming to be the smartest. One is excellent at coding, another shines in reasoning, while another performs better in scientific questions or creative writing.

For Non-Members: Read here!

But what if we stop asking, Which is the best AI model? and instead start asking, What if the best answer comes from multiple models working together?

That is exactly the idea behind Sakana Fugu, a new family of orchestrator models developed by Sakana AI. We build several powerful models working together like a skilled team instead of focusing on building one giant model.

Let’s understand this better!

The Problem with a Single AI Model

Imagine you have three friends.

• One is an excellent software engineer, and he helps you when your laptop crashes.
• One is a mathematician, and he helps you when you want to solve a difficult equation.
• One is a scientist, and he helps you when you are working on a biology experiment.

Now imagine asking only the mathematician to repair your laptop. It probably won’t end well.

Today’s frontier language models are surprisingly similar. Each one has unique strengths. Some are exceptional at software engineering, and some are great at mathematical reasoning. And some perform better in science or long-form reasoning.

Instead of forcing one model to do every task, Sakana AI believes the smarter solution is to let each model contribute where it performs best.

This sounds obvious when humans do it. Yet until recently, AI systems mostly relied on a single model for everything.

What is Sakana Fugu?

Sakana Fugu is not another large language model trying to replace GPT or Claude.

It is an orchestrator; you can think of it like a project manager. When you ask a question, Fugu first understands the problem. Then it decides:

• Which AI model should work on this?
• Should multiple models participate?
• Should they work together?
• Should one model verify another’s answer?
• How should all responses be combined?

The user simply receives one final answer. But behind the scenes, several models may have collaborated to produce it. This is why Sakana calls it collective intelligence.

Two Versions, Two Different Goals

The Fugu family contains two models.

1. Fugu

This version of the model focuses on speed. It quickly analyzes the question and routes it to the single worker model that is most likely to solve it well.

This keeps the response time close to calling one frontier model directly while still choosing the best expert for each task. You can think of it like asking the right expert immediately.

2. Fugu-Ultra

This model takes a different approach; instead of asking only one expert, it builds an entire workflow. It may assign different subtasks to different models.

• One model plans.
• Another writes code.
• A third checks mistakes.
• A fourth combines everything into the final response.

Naturally, this takes more time, but it often produces higher-quality answers for difficult problems.

An Orchestra Instead of a Solo Singer

The name orchestrator reminds me of an orchestra. An orchestra does not ask the violin to play every instrument.

• The violin plays its part.
• The piano plays its own.
• The drums add rhythm.

Finally, the conductor makes sure everyone performs together beautifully. Fugu-Ultra behaves exactly like that conductor.

It creates a custom workflow for every question instead of using the same process every time. That flexibility is one of its biggest strengths.

How Does Fugu Learn?

How does Fugu know which model to choose? The answer is interesting.

Step 1: Learn from many examples

The developers first collect thousands of problems covering coding, reasoning, mathematics, science, and many other areas. Every worker model solved every problem. The system measured which model consistently performed best.

Instead of learning only one correct model for each question, Fugu learns a probability distribution showing which models usually perform well on different kinds of tasks.

Step 2: Learn from real workflows

After that, the training becomes more realistic. Rather than solving isolated questions, Fugu learned from complete software engineering workflows.

These included editing files, calling tools, fixing bugs, running code, and reacting to execution errors. This helped Fugu understand which models perform well inside real-world environments instead of only scoring highly on benchmarks.

I think this is a very practical direction because most AI applications today involve much more than answering a single question.

What Makes Fugu-Ultra Different?

Fugu-Ultra goes one step further; instead of selecting one worker, it actually designs a workflow. For example, it may produce something like this:

1. Model A studies the problem.
2. Model B writes an initial solution.
3. Model C reviews it carefully.
4. Model A combines everything into the final answer.

The impressive part is that this workflow changes depending on the user’s question. No fixed template is used. Every task gets its own plan.

Why Memory Matters

Imagine three developers are working on the same software project. If all three constantly overwrite each other’s work, the project becomes chaotic. If they never communicate, important information gets lost.

Fugu-Ultra solves this by separating working memory. Each agent keeps its own reasoning process while sharing only the information that is actually useful for collaboration.

This prevents one model from simply copying another while still allowing effective teamwork.

Performance That Caught My Attention

The technical report compares Fugu with several leading frontier models across coding, reasoning, science, and agent benchmarks.

According to Sakana AI, Fugu-Ultra achieves state-of-the-art or near state-of-the-art performance on several difficult evaluations, including:

• SWE-Bench Pro
• Terminal Bench
• LiveCodeBench
• GPQA Diamond
• Humanity’s Last Exam
• CharXiv Reasoning

The important point isn’t just the benchmark scores, but the more interesting observation is how those scores are achieved.

Fugu is not becoming smarter by increasing model size. Instead, it becomes smarter by coordinating multiple specialized models effectively.

Real Examples of Intelligent Coordination

Several fascinating examples include:

• During coding tasks, one model often handles the implementation while another steps in specifically to debug mistakes.
• For mathematical reasoning, GPT-style models are frequently chosen for computation-heavy work.
• For scientific questions, another model with stronger scientific ability was selected.
• Sometimes one model acts purely as a reviewer, checking another model’s work before producing the final answer.

This feels very similar to how experienced engineering teams work. Everyone has a specialty, and each contributes differently.

Why This Matters for the Future of AI Agents

Personally, I think Sakana Fugu represents something larger than another benchmark improvement.

Today’s AI discussions often revolve around building bigger models, more GPUs, more parameters, longer context windows. But Fugu suggests another direction. Instead of making one model infinitely larger, perhaps we should build better teams of models.

This has several advantages.

• New models can be added without retraining the whole system.
• Organizations can choose which providers they trust.
• Companies can replace one worker model without redesigning everything.

The system becomes modular rather than monolithic.

My Final Thoughts

Reading the Sakana Fugu report changes the way we think about AI systems. For years, the industry has focused on creating the single smartest model.

Sakana AI asks a different question.

What if intelligence comes from collaboration instead of competition?

That idea feels surprisingly natural. Humans have always solved difficult problems by working together. Why shouldn’t AI work the same way?

I believe the orchestrator models like Fugu will become increasingly important as AI agents continue to evolve. Future AI systems may not be judged by how powerful one model is, but by how intelligently they coordinate many specialized models.

And perhaps that is the next chapter of artificial intelligence, not bigger brains, but better teamwork.

Digital Products

ML Interview Book: Crack Your Next ML Interview with Machine Learning Interview Playbook

Productivity Tool: ***Social Media Time Tracker: Take Back Your Time, a tool that annoys you when you log in to social media sites. Chrome Extension.***

Connect with the author

LinkedIn | YouTube | Threads | Instagram | Facebook

Reference

[embed]