Sakana AI

Sakana AI is an AI research and development company headquartered in Tokyo, Japan. It was founded in July 2023 by David Ha, Llion Jones and Ren Ito; as of October 2026 Ha is CEO, Jones is CTO and Ito is chairman. Ha led the Google Brain research team in Tokyo, and Jones is a co-author of the 2017 paper ‘Attention Is All You Need’, which introduced the Transformer architecture behind today's large language models. Ito worked at Japan's Ministry of Foreign Affairs, Mercari and Stability AI. The company's logo is inspired by schools of fish, reflecting its aim to bring ‘collective intelligence’ — many individuals moving together — into AI development.

Sakana AI is known for research that combines existing models efficiently rather than building huge models from scratch. In March 2024 it announced ‘Evolutionary Model Merge’, which imitates biological evolution to merge several open models automatically, and released models such as the Japanese maths model EvoLLM-JP. The work was published in the journal Nature Machine Intelligence in January 2025. In August 2024 it announced ‘The AI Scientist’, which carries out the whole research process on its own, from ideas to experiments and paper writing (under $15 per paper, developer-reported), and this work was published in Nature in March 2026. In April 2026 it opened a beta of Sakana Fugu, a service that picks suitable AI models for each problem and has them work together.

Its Series B round was announced in November 2025 at about ¥20 billion; in an April 2026 update the company put the total at about ¥32 billion (about $200 million) and the post-money valuation at about ¥432 billion (company-reported). Investors include Mitsubishi UFJ Financial Group (MUFG) and Google, and it builds AI for finance together with MUFG and Daiwa Securities Group. On 29 July 2026 it signed a contract with Japan's Ministry of Defense to research and demonstrate the AI capabilities needed for integrated analysis work. On 26 October 2026 it announced that it will build a Japanese language model, provisionally named ‘Sawara’, itself starting from pre-training (the first stage, in which a model reads huge amounts of text to gain basic abilities). The company describes this as part of its effort towards ‘AI sovereignty’, so that it does not rely only on foreign models.