This $2.6B Japanese A.I. Startup Exposes the Language Gap in Today’s LLMs

Like many other A.I. startups, Sakana AI is pursuing technology that can reason, self-improve and aid scientific research. But where the Tokyo-based venture stands apart from most of the world’s model developers is in its focus on designing products specifically for Japanese society, language and industries.

Investors are thrilled about the idea. Sakana announced today (Nov. 17) that it had raised around $135 million in a Series B round that values the company at $2.6 billion, making it one of Japan’s most valued A.I. startups. Backers include Mitsubishi UFJ Financial Group (MUFG), Khosla Ventures, NEA and Lux Capital, among others.

Sakana was founded in 2023 by a group of former Google and Stability AI researchers, including CEO David Ha. The young company’s products so far include reasoning models, a self-improving coding agent and “AI Scientist,” a system designed for scientific discovery.

The startup also works with major Japanese institutions to develop customized products for the finance industry. Its clients include MUFG, Japan’s largest bank, and Daiwa Securities Group. Sakana is now eyeing expansion into defense, intelligence and manufacturing.

“As a business leader myself, I feel a responsibility to lead the implementation of A.I. across all Japanese companies,” said Hironori Kamezawa, CEO of MUFG, in a statement. “Through this investment, I hope the benefits of A.I. will extend beyond transforming banking operations to Japan’s diverse industries.”

Building sovereign A.I.

Alongside expansion plans and potential investments, partnerships and potential acquisitions, Sakana will use its new capital to advance research into large language models (LLMs) that reflect Japan’s national culture and values. “This allows us to sustainably develop and optimize frontier models for Japan’s specific needs, offering a distinct path and resource requirement from the large-scale model competition,” the company said in its funding announcement.

More than half of all online content today is written in English, which poses a challenge for training LLMs that serve non-English markets. Fewer than 5 percent of the approximately 7,000 languages spoken around the globe have meaningful representation online, according to a recent paper from the Stanford Institute for Human-Centered AI that described LLMs as suffering from a “digital divide.”

Sakana is not alone in attempting to close this gap. Krutrim, the A.I. startup launched by Ola co-founder Bhavish Aggarwal, is building LLMs in Indian languages, while France’s Mistral AI has positioned its Le Chat assistant as a European alternative to OpenAI’s ChatGPT. Earlier this year, Microsoft announced a slew of initiatives, including a call for proposals to expand training data in European languages, to spur more multilingual A.I. development.

Despite the limited number of A.I. systems tailored for Japan, Sakana argues the demand and benefits are clear. “For Japan, with a declining workforce and an aging population, the benefits of A.I. technology are clear,” said the startup. “We believe the right path for Japan, a nation with limited resources, is to develop our A.I. ecosystem sustainably through innovation and collaboration, with a strong focus on societal implementation of the technology.”