Sakana AI

Sakana AI is a Japanese artificial intelligence company founded in Tokyo in 2023 by former Google researchers David Ha, Llion Jones (a co-author of the Transformer paper), and Masato Hagiwara. The company develops AI models using nature-inspired evolutionary and collective-learning methods, and has become one of Japan's highest-valued AI startups, reaching a valuation of approximately US$2.65 billion after its November 2025 Series B round.^[1]

History

Sakana AI was established in September 2023 amid a wave of Japanese startups aiming to build domestic alternatives to American and Chinese AI labs. Its name, meaning "fish" in Japanese, reflects the founders' framing of AI training as an evolutionary, ecosystem-like process rather than a single monolithic model.^[2]

The company raised about US$200 million in a Series A round (announced 2024) at a reported US$1.5 billion valuation, with investors including Khosla Ventures, Lux Capital, and New Enterprise Associates. In November 2025 it closed a ¥20 billion (≈US$135 million) Series B at a US$2.65 billion post-money valuation, bringing total funding to roughly US$350–380 million — one of the largest Japanese AI fundraises of that year.^[1][3]

Technology and research

Sakana AI's stated research direction departs from the dominant "scale everything" paradigm:

• Evolutionary model merging. Rather than training one large model from scratch, Sakana combines existing trained models using evolutionary algorithms, selecting and recombining candidate models against a fitness objective — an approach the company branded model fusion.
• Continuous learning / lifelong learning. The lab has published work on agents that keep learning from new data without forgetting old knowledge, reducing the cost of repeatedly retraining foundation models.
• Japanese-language optimisation. Its consumer and enterprise models are tuned for Japanese text, code, and cultural context — a market where Western models have historically weaker coverage.
• The Transformer paper lineage. Co-founder Llion Jones's presence gives the company a direct connection to the architecture underlying virtually all modern large language models.^[2]

Sakana also gained attention for an AI Scientist concept — an autonomous pipeline for generating and running machine-learning experiments — and for an early dispute in 2024 in which the company acknowledged its then-recent PhD programme proposal contained overlapping passages with other submissions, revising its processes afterward.

Market position

Sakana is often described as "Japan's answer to OpenAI," though its strategy is narrower: specialised, efficient models for the Japanese and broader Asian market, using training innovations to compete on cost rather than raw scale. It sits alongside Asia's other sovereign-minded AI labs — China's Moonshot AI, Zhipu AI, and Alibaba's Qwen, and Southeast Asia's SEA-LION — as evidence that regional language and data priorities are spawning local champions rather than leaving the field to US labs.

>Key Takeaways

• Sakana AI is a Tokyo lab (founded 2023) by ex-Google researchers including Transformer co-author Llion Jones.
• It competes through nature-inspired training — evolutionary model merging and continual learning — rather than pure scale.
• It reached a US$2.65B valuation in November 2025, making it one of Japan's top AI startups.
• Its Japanese-language focus is a template for Malaysia's own sovereign Malay-language model ambitions.

See Also

• Moonshot AI
• Zhipu AI
• Qwen
• Sea-LION
• ILMU Malaysian LLM
• Model Merging