What is Sakana.ai?

What is Sakana AI and what does it actually do?

Sakana AI is an artificial intelligence company headquartered in Tokyo, Japan. The name “Sakana” means “fish” in Japanese, and the company draws inspiration from how schools of fish move together using collective intelligence. Instead of building one giant model and calling it a day, Sakana AI explores how many smaller, specialized systems can work together to solve complex problems.

The company does two things at once. First, it publishes cutting-edge research on topics like multi-agent orchestration, recursive self-improvement, and automated scientific discovery. Second, it ships real products that people can use today, including chat tools, translation services, research agents, and enterprise APIs. This dual focus on research and practical products is what sets Sakana AI apart from many labs that only publish papers.

If you are curious about how other AI tools compare, you can read our reviews of Meshy AI, Inworld AI, and Tripo 3D to see how different platforms approach generative AI.

Who founded Sakana AI and why does the leadership matter?

Sakana AI was founded in July 2023 by three people: David Ha, Llion Jones, and Ren Ito. David Ha serves as CEO and previously led the Japan Research Team at Google Brain. He is well known for his work on complex self-organizing systems and the evolution of artificial intelligence. Llion Jones serves as CTO and is one of the co-authors of the famous “Attention Is All You Need” paper that introduced the Transformer architecture. That paper changed the entire AI industry and sits underneath almost every modern language model. Ren Ito serves as Chairman and brings deep experience from leadership roles at Mercari and Stability AI, plus a background in law and Japan’s Ministry of Foreign Affairs.

This leadership team matters because it combines world-class research credibility with serious business and policy experience. When a company is run by someone who co-wrote the paper that made ChatGPT possible, people pay attention. When that same company also has executives who understand Japanese enterprise needs and government relations, it becomes easier to turn research into products that actually get adopted.

What products does Sakana AI offer in 2026?

In 2026, Sakana AI moved fast from research to product delivery. The company now offers a growing family of tools that cover chat, translation, deep research, and enterprise orchestration. Here is a quick breakdown of the main products you should know about.

For readers interested in how AI automation tools fit into broader workflows, our guide to Gumloop AI automation explains how no-code AI pipelines are changing how teams work.

What is Sakana Namazu and why is it special for Japanese speakers?

Sakana Namazu is the company’s flagship language model designed specifically for Japanese users. It is built on top of the open Kimi K2.6 model and then fine-tuned with Sakana AI’s own in-house data focused on Japanese language, business customs, and cultural context. The result is a model that does not just speak Japanese, but actually understands how Japanese people communicate at work and online.

Most global AI models treat Japanese as just another language to support. Namazu treats it as a first-class citizen. It handles business honorifics (keigo), catches subtle cultural references, and understands internet slang that generic models often miss. It also supports web search and code execution, so it can handle complex tasks rather than just answering simple questions.

The model offers a 262,000-token context window, which means you can feed it long documents, entire codebases, or lengthy conversations without losing coherence. For developers, it is available through an OpenAI-compatible API, making it easy to drop into existing applications.

How does Sakana Fugu work and what is multi-agent orchestration?

Sakana Fugu is probably the most technically interesting product in the lineup. Instead of relying on a single AI model to do everything, Fugu acts like a conductor in an orchestra. It dynamically assembles a team of specialized models, assigns them roles, and coordinates them to solve complex tasks.

Here is a simple way to think about it. Imagine you have a really hard coding problem. One model might be great at writing Python, another at debugging, and another at checking security. Fugu figures out which models to call, in what order, and how to combine their outputs into one solid answer. It does this through learned coordination strategies rather than hard-coded rules, which means it gets better at teamwork over time.

Fugu comes in three flavors. Fugu is the balanced option for everyday work. Fugu Ultra pulls in a deeper pool of expert agents for high-stakes problems like Kaggle competitions, cybersecurity analysis, and patent research. Fugu Cyber is specialized for security tasks like vulnerability research and threat investigation. All three are accessible through a single API endpoint, so you do not need to rewrite your integration when switching between them.

If you are building with AI-powered 3D tools, you may also want to check out our coverage of Hyper3D Rodin AI and Sloyd AI for a broader view of the generative AI landscape.

What is Sakana Marlin and who is it built for?

Sakana Marlin is an enterprise research agent marketed as a “Virtual CSO” or Virtual Chief Strategy Officer. It is designed for businesses that need deep, thorough research but do not have weeks to spare. Once you give Marlin a research topic, it can run continuously for up to 8 hours, reasoning through sources, synthesizing findings, and generating reports that can exceed 100 pages with slides included.

Early users have applied Marlin to competitive intelligence, market entry analysis, literature reviews, and strategic planning. The idea is to let AI handle the heavy lifting of research while human executives focus on decision-making. This is a different approach from standard chatbots that give instant but shallow answers. Marlin trades speed for depth, which is exactly what serious business research often requires.

What makes Sakana Translate different from Google Translate?

Sakana Translate is a Japanese-English-Chinese translation tool that runs on the Namazu model. The big difference from generic translation services is that it was built by people who live and work in Japan and understand where standard tools fall short. It includes Translate, Proofread, and Ask modes, so you can not only convert text but also refine it or ask follow-up questions about nuance.

Standard translation engines often struggle with Japanese business emails, honorific levels, and culturally specific phrases. Sakana Translate was designed specifically to handle these pain points. David Ha, the company’s CEO, has said publicly that he uses the tool every day, which suggests the team is dogfooding their own product rather than shipping something they do not personally rely on.

What research breakthroughs has Sakana AI published?

Sakana AI is not just a product company. Its research team has published several notable projects that have attracted attention across the AI community. Some of the most important include:

• The AI Scientist — An early project exploring how AI systems can automate parts of the scientific research process.
• Darwin Gödel Machine — A system focused on recursive self-improvement, where AI agents learn to improve themselves over time.
• TRINITY and The Conductor — Two research papers presented at ICLR 2026 that form the technical foundation of Fugu. They show how systems can learn to assemble, route, and coordinate expert agents without hand-designed workflows.
• AB-MCTS — A method for autonomous research that powers the deep reasoning capabilities inside Sakana Marlin.

This research output gives Sakana AI credibility in academic and technical circles. It also means their products are not just wrappers around someone else’s API. There is genuine original research underneath the commercial offerings.

For those interested in AI safety and responsible development, our article on Roblox open-sourcing AI safety tools covers how the industry is thinking about keeping AI systems safe as they grow more powerful.

How is Sakana AI different from OpenAI, Google, and Anthropic?

The biggest difference is focus. OpenAI, Google, and Anthropic are building general-purpose frontier models for the entire world. Sakana AI is building frontier AI specifically for Japan, while also contributing to global research. This local focus shows up in everything from the training data to the product design to the enterprise partnerships.

Another difference is philosophy. Sakana AI believes in nature-inspired, collective intelligence rather than brute-force scaling of single monolithic models. While US labs often compete on who has the biggest model, Sakana asks whether a swarm of smaller, specialized models working together might be more efficient and more adaptable.

Finally, Sakana AI has deep ties to Japanese enterprises, government, and public sector organizations. Its investor list includes major Japanese banks, insurers, and industrial giants like NEC, Fujitsu, and Mitsubishi Electric, alongside international names like NVIDIA, Google, and Khosla Ventures. In late 2025, the company raised a Series B at a valuation of roughly $2.6 billion, making it one of Japan’s largest startups ever.

Frequently Asked Questions About Sakana AI