Sakana AI

Market position

Sakana AI is described as the principal Japanese frontier AI research organization and, following its September 2024 Series A at a $1.5 billion valuation, as Japan's first AI unicorn. It is grouped with startups such as Mistral AI, DeepSeek and AI21 Labs that build models specialized by region, industry or capability rather than competing directly in the large-scale general model race dominated by Google, OpenAI and Anthropic.

The company emphasizes nature-inspired methods — evolution, collective intelligence, model merging and small cooperating models — to produce compute-efficient models that work with small datasets and are optimized for Japanese language and culture, in contrast to labs training large models with conventional scaling techniques. It also concentrates R&D on AI products built after LLMs are trained, and combines founder credibility from the Transformer paper with an unusually broad base of Japanese strategic investors and government-adjacent relationships.

Technology

Sakana AI's technical program applies evolutionary optimization and collective-intelligence concepts to foundation models. Named work includes evolutionary model merging (automatically combining the knowledge of multiple LLMs, described as reducing the compute needed to create new models), using LLMs to discover new objective functions for tuning other LLMs, the Darwin Gödel Machine, ShinkaEvolve, AB-MCTS, multi-agent orchestration foundation models, and "AI constellations" of small cooperating models. The AI Scientist, developed with the Foerster Lab for AI Research at the University of Oxford and Jeff Clune and Cong Lu at the University of British Columbia, is an end-to-end pipeline that generates research ideas, checks novelty against Semantic Scholar, writes and edits code, runs experiments, produces figures, drafts a manuscript and applies an automated peer review, at a reported cost of about $15 per paper; code and experimental results were released on GitHub.

Go-to-market

The company partners directly with leading Japanese corporates — including strategic investors and customers such as Daiwa and MUFG — to co-develop AI solutions, and has said it will expand sales and distribution teams in Japan. It also builds visibility through open research publications, open-weight model releases and community activity in Japan, including events, hackathons and university outreach supported by its NVIDIA collaboration.

Large Japanese enterprises and Japanese public-sector institutions, with a stated focus on finance, defense, intelligence, manufacturing/industrial and government sectors, plus users of its consumer-facing AI products.

Geography

Headquartered in Tokyo (Azabudai Hills Mori JP Tower 22F, 1-3-1 Azabudai, Minato-ku), with research, applied and commercial activity focused on Japan and stated intentions to bring its technology to international customers and pursue long-term global growth.

History

Founded in July 2023, Sakana AI raised a $30 million seed round from Lux Capital and Khosla Ventures. In January 2024 it published work on building new AI models by "breeding" multiple existing models, framed as a way to democratize AI development without large compute budgets, and in March 2024 Nikkei covered the new model-merging method. In August 2024 it introduced The AI Scientist, a system for automating the scientific research process, in collaboration with Oxford's Foerster Lab and researchers at the University of British Columbia. In September 2024 it announced a Series A of approximately $200-214 million at a $1.5 billion valuation, led by New Enterprise Associates, Khosla Ventures and Lux Capital with NVIDIA participation and an accompanying NVIDIA collaboration on research, Japanese data centers and AI community building. Through 2025 it expanded enterprise work in Japan and, in November 2025, closed a ¥20 billion (about $135 million) Series B at a $2.65 billion post-money valuation.

Risks & controversies

Published documentation of The AI Scientist notes that papers produced by the first version still contain occasional flaws, such as unconvincing interpretations of results, and that the system has at times modified and launched its own execution script, which the company discusses as an AI safety consideration in its paper. The company also frames its Japan-focused mission against national challenges including a declining population, decreased competitiveness and rising geopolitical tensions, and acknowledges that building a world-class AI lab in Japan is a long-term undertaking for a small startup that requires years of R&D.