Sakana AI: nature-inspired AI from Tokyo

A lab founded by renowned researchers

Sakana AI was founded in July 2023 in Tokyo. Its founders include Llion Jones, one of the authors of the 2017 paper “Attention Is All You Need”, which introduced the Transformer architecture behind almost every modern language model. David Ha, the other research co-founder, runs the company. Its headquarters are in Tokyo’s Azabudai district.

Sakana describes itself as a frontier AI research and development company that works with Japanese businesses and the country’s public sector.

Research inspired by nature

What sets Sakana apart is its approach. Rather than building ever-larger models, the lab explores evolution and collective intelligence. Some notable work:

• evolutionary model merging: combining several existing models, much like breeding, to produce a new one (EvoLLM-JP, EvoVLM-JP);
• The AI Scientist, a system that tries to automate an entire research project, from idea to paper;
• the Darwin Gödel Machine, an agent that rewrites and improves its own code, and ShinkaEvolve, which evolves programs with the help of language models;
• work on orchestrating several cooperating models, close to multi-agent systems.

Namazu: models built for Japanese

On the product side, Sakana offers Namazu, a family of large language models specialized in Japanese and in the norms of Japanese business. The lab does not train them from scratch: it starts from existing open-weight models, such as Kimi K2.6 from Moonshot AI, then adapts them with its own data. Since August 2026, Namazu has been available through an OpenAI-compatible API, with built-in web search and code execution.

Its openness policy is therefore mixed: research models are often published on Hugging Face, while Namazu is offered as a service. Sakana’s product page does not say whether Namazu’s weights can be downloaded.

How to get access

For Namazu, create a key in the Sakana AI console or go through OpenRouter. The research models live on Hugging Face under the SakanaAI organization. To compare these models with others, use our model comparison tool.