Sakana AI – What It Is, How It Works, and Why It Could Change the Future of AI

Sakana AI is one of the fastest-growing AI companies in the world, having become a Japanese unicorn valued at over $2.6 billion in just a few years (Startup Intros – Sakana AI Funding). But this isn’t just another startup churning out another language model. Sakana AI focuses on collective intelligence, evolution, and localized adaptation—what it calls “AI that creates AI.”

In this article, we’ll explore:

• what Sakana AI is and when it was founded,
• how its core technologies—Namazu, Fugu, and Marlin—work,
• what problems they solve (e.g., refusal to answer, export restrictions, long-term research),
• who’s behind it and why it could matter for the future of AI.

---

---

1. What Is Sakana AI and When Was It Founded?

Sakana AI is a Tokyo-based research and development lab focused on Frontier AI—technology that meets the highest global standards. The company was founded in July 2023 by three experienced researchers and managers:

• David Ha – CEO, former Head of Research at Google Brain and Head of Research at Stability AI.
• Llion Jones – CTO, co-author of the groundbreaking paper “Attention Is All You Need”, which introduced the Transformer architecture.
• Ren Ito – Chairman, former COO of Stability AI with experience scaling AI startups.

Its official mission is: “Building Frontier AI in Japan”—developing world-class AI tailored to Japan’s needs while competing with global giants (Sakana AI Corporate Info).

The company operates across three main pillars:

1. Research – nature-inspired research (evolution, collective intelligence),
2. Applied – solutions for finance, defense, and infrastructure,
3. Product – tools such as Sakana Chat, Sakana Marlin, and Sakana Fugu.

Sakana AI quickly secured funding from top-tier VC firms (including Khosla Ventures, Lux Capital, and Nvidia) and major Japanese megabanks (MUFG, SMBC, Mizuho), becoming one of Japan’s fastest-growing unicorns (Startup Intros – Sakana AI Funding).

---

2. Namazu – How to Turn a Global Model Into a “Japanese” One Without Losing Power

2.1. The Problem: Global Models Don’t Fit Japan

Most large language models (LLMs) come from the U.S. or China. They’re trained on global data, but:

• they may carry strong ideological biases (e.g., political, historical),
• they often refuse to answer on sensitive topics (self-censorship),
• they’re not optimized for the Japanese language or local cultural context.

According to Sakana AI’s research, some foreign models refuse to answer questions about politics, history, or diplomacy in 72% of cases (BigGo Finance – Namazu post-training). This makes them practically useless in business or administration, where objective, fact-based responses are required.

2.2. The Solution: Namazu Post-Training

Sakana AI developed a post-training technique that:

• takes existing large open-source models (e.g., DeepSeek-V3.1-Terminus, Llama-3.1-405B),
• and adapts them to Japan using specialized training on data that reflects Japanese cultural and security contexts.

The result is a series of prototype models called Namazu (alpha version) that:

• maintain high performance on standard benchmarks (AIME’25, MMLU-Redux, GPQA Diamond, LiveCodeBench, IFEval),
• while radically improving neutrality and accuracy on political and historical topics (Sakana AI – Namazu Alpha).

Key outcome: the refusal rate for sensitive questions dropped from 72% to nearly 0% in the Namazu-DeepSeek-V3.1-Terminus model (BigGo Finance – Namazu post-training).

2.3. How Does It Work Technically?

Namazu isn’t a model built from scratch. It’s an adaptation of existing models:

1. A high-quality open-source model is selected (e.g., DeepSeek-V3.1-Terminus).
2. Sakana AI builds a specialized dataset that reflects Japanese cultural, political, and security contexts.
3. The model is fine-tuned (post-trained) to:
   • better understand Japanese,
   • be more neutral in responses,
   • stop refusing to answer sensitive topics without justification.

Benchmark results show that Namazu performs comparably to base models on math, logic, and coding tasks, but significantly outperforms them on politics and history (Sakana AI – Namazu Alpha).

2.4. Sakana Chat – A Free Namazu-Powered Chatbot

Built on Namazu is Sakana Chat—a free chatbot primarily for users in Japan that:

• uses Namazu models as its engine,
• includes web search capabilities,
• allows real-time comparison and integration of information from multiple sources (GIGAZINE – Sakana Chat & Namazu).

Sakana Chat was tested by around 1,000 beta users, and their feedback helped refine both the model and the interface (Sakana AI – Namazu Alpha).

---

3. Sakana Fugu – A Multi-Agent System as One Model

3.1. Why One Model Isn’t Enough

Most AI companies offer single models: GPT-5.5, Claude Opus, Gemini, etc. But every model has strengths and weaknesses. Sakana AI took a different path: instead of building one giant model, it created a multi-agent system that behaves like a single model.

Sakana Fugu is:

• a multi-agent system (MAS),
• that dynamically coordinates a pool of different LLM models,
• accessible via a single OpenAI-compatible API (Sakana AI – Fugu).

From the user’s perspective, it works like this:

• you send a query to one endpoint,
• you specify a model fugu or fugu‑ultra,
• and Fugu internally decides which models to use and how to coordinate them.

3.2. Architecture: TRINITY and Conductor

Fugu’s architecture is based on two ICLR 2026 papers:

• TRINITY – a lightweight coordinator that assigns roles to agents:
  • Thinker – plans,
  • Worker – executes tasks,
  • Verifier – validates results.
• Conductor – a model trained with RL to discover coordination strategies in natural language (Sakana AI – Fugu).

In practice, this means Fugu can:

• solve simple tasks on its own,
• or assemble a team of experts (different models) and coordinate their work,
• while the user sees only one consolidated response.

3.3. Fugu vs. Fugu Ultra

Sakana Fugu offers two variants:

1. Fugu – a balance between performance and latency, designed for everyday use. It allows excluding specific agents from the pool (e.g., if you don’t want to use a particular provider).
2. Fugu Ultra – optimized for maximum response quality on complex tasks (e.g., Kaggle competitions, cybersecurity analysis), using a deeper pool of experts (fixed pool, no exclusion option) (Sakana AI – Fugu).

3.4. Benchmarks: Fugu Ultra vs. GPT-5.5, Opus, and Gemini

According to benchmarks shared by Sakana AI, Fugu Ultra achieves comparable or better results than leading models:

• SWE Bench Pro (software engineering problem-solving):
  • Fugu Ultra: 73.7%,
  • GPT-5.5: 58.6%,
  • Gemini 3.1 Pro: 54.2% (Sakana AI – Fugu).
• TerminalBench 2.1 (agentic coding): Fugu Ultra: 82.1.
• LiveCodeBench Pro: Fugu Ultra: 90.8.
• GPQA-D (scientific knowledge): Fugu Ultra: 95.5.

This means Fugu doesn’t just combine multiple models—it outperforms them as a system.

3.5. How to Use the Fugu API?

Fugu is available via an OpenAI-compatible API:

• Base URL: https://api.sakana.ai
• Model ID: fugu or fugu‑ultra
• API key generated in the console: https://console.sakana.ai (Sakana AI – Fugu).

Example in Python:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.sakana.ai",
    api_key="sk-...",  # Twój klucz z console.sakana.ai
)

response = client.chat.completions.create(
    model="fugu",  # lub "fugu-ultra"
    messages=[
        {"role": "user", "content": "Wyjaśnij, czym jest Sakana Fugu."}
    ],
)

print(response.choices[0].message.content)

This makes it easy to replace OpenAI GPT with Fugu in existing applications.

---

4. Sakana Marlin – A Virtual CSO for Ultra-Deep Research

4.1. What Is Marlin?

Sakana Marlin is Sakana AI’s first commercial product that isn’t a language model—it’s an autonomous research agent. Described as “Your Virtual CSO (Chief Strategy Officer)”, it’s a tool for ultra-deep strategic research (Sakana AI – Marlin).

While most AI tools (e.g., ChatGPT Deep Research, Gemini Deep Research) focus on speed (responses in seconds), Marlin deliberately slows down the process:

• it operates for up to 8 hours of continuous, autonomous reasoning,
• generates reports ranging from dozens to ~100 pages plus slides for executives,
• is designed for corporations, financial institutions, think tanks, and consultants (VentureBeat – Sakana Marlin).

4.2. How Does It Work Technically? AB-MCTS and Multi-Model Collaboration

Marlin is built on the AB-MCTS (Adaptive Branching Monte Carlo Tree Search) architecture—a method that enables AI to perform trial-and-error and explore multiple hypotheses simultaneously (Sakana AI – AB-MCTS).

Combined with multiple LLM models (e.g., OpenAI o4-mini, Gemini 2.5 Pro, DeepSeek R1-0528), Marlin:

• formulates hypotheses,
• scours the web,
• resolves contradictions between sources,
• and delivers exhaustive, expert-level strategic reports.

It’s a product directly rooted in Sakana AI’s earlier research, such as The AI Scientist (a system that automatically generates academic papers) (Sakana AI – The AI Scientist).

4.3. Business Applications

Marlin is designed for tasks like:

• market and competitor analysis,
• assessing technology trends,
• regulatory scenarios (e.g., stablecoins, payment tokenization),
• AI agent market maps for enterprises (Sakana AI – Marlin).

Example topics Marlin can research:

• “Scenarios for a blockade of the Strait of Hormuz and their impact on the global economy”,
• “Impact of stablecoin regulations on payment systems”,
• “AI agent market map for large corporations”.

Marlin doesn’t replace human decision-making, but it dramatically reduces the time needed for research, allowing teams to focus on strategic choices.

---

5. Who’s Behind Sakana AI and Why It Matters

5.1. The Founders and Their Expertise

• David Ha – CEO, known for research in generative models and evolutionary algorithms. Previously led teams at Google Brain and Stability AI.
• Llion Jones – CTO, co-author of “Attention Is All You Need,” the foundational paper behind modern LLMs.
• Ren Ito – Chairman, with experience scaling AI startups and global operations.

Their blend of academic and business experience makes Sakana AI both a research lab and a product company.

5.2. Mission: AI for Japan, with Global Reach

Sakana AI isn’t aiming to be just a local provider. Its goal is to:

• build sovereign AI for Japan—solutions not entirely dependent on foreign suppliers,
• while competing with global companies on the international stage.

In an interview with Science Japan, Ren Ito said the key is thinking not in terms of “foreign vs. Japan,” but “U.S. West Coast vs. the rest of the world”—and building a world-class company whose technology happens to be based in Japan (Science Japan – Interview with Ren Ito).

5.3. Why Sakana AI Could Change the Future of AI

Sakana AI stands out in several key ways:

1. Post-training over giant pretraining – instead of spending hundreds of millions on training from scratch, it adapts existing models to local needs. This is more scalable and cost-effective.
2. Multi-agent systems as a product – Fugu shows that the future of AI may not lie in bigger models, but in better coordination of many models.
3. Long-term reasoning – Marlin proves AI can operate not just in seconds, but in hours, opening the door to more complex tasks.
4. Focus on Japan – while most AI companies target the U.S./EU, Sakana AI shows that regional needs can drive innovation.

---

6. Summary: What Sakana AI Means for Users and Developers

For end users:

• Sakana Chat offers a free, Japan-optimized chatbot with web search.
• Sakana Marlin is a tool for deep strategic research for businesses.
• Sakana Fugu is an API that can replace traditional models in applications, often delivering better quality thanks to multi-agent coordination.

For developers:

• Fugu provides an OpenAI-compatible API that can be easily integrated into existing tools.
• Namazu demonstrates how to adapt global models to local needs without sacrificing performance.
• The multi-agent architecture of Fugu and Marlin’s long-term reasoning could inspire new approaches to AI system design.

Sakana AI is more than just another AI startup. It’s an example of how collective intelligence, evolution, and localized adaptation can change the way we build and use artificial intelligence. If you want to follow their progress, check out:

• Official website: https://sakana.ai
• Blog: https://sakana.ai/blog
• Sakana Chat: https://chat.sakana.ai
• Sakana Fugu: https://sakana.ai/fugu
• Sakana Marlin: https://sakana.ai/marlin