ELYZA Speaks Japanese

Private

ELYZA is a Tokyo AI company that grew out of the University of Tokyo’s Matsuo Lab and builds large language models tuned for Japanese. In 2024 KDDI, one of Japan’s three big mobile carriers, took majority control, turning ELYZA into the group’s in-house generative AI team. This profile looks at how the company works, what it ships, and why it matters to overseas firms doing business in Japan.

From research lab to enterprise vendor

ELYZA was founded in 2018 by researchers connected to the Matsuo Lab, the University of Tokyo group led by Professor Yutaka Matsuo that has produced a long line of Japanese AI startups. The company began with natural language processing work for large Japanese enterprises: summarising documents, drafting text and automating back-office reading tasks that are hard to hand to English-first tools.

That early focus shaped its strategy. Rather than racing to train the largest possible model from scratch, ELYZA concentrated on making strong open models perform well in Japanese, then wrapping them in products that companies can actually deploy.

The model strategy: open base, Japanese training

ELYZA’s best-known releases take a public base model and add extensive Japanese-language training on top.

• ELYZA-japanese-Llama-2 — released in 2023, built on Meta’s Llama 2 and published openly, it became one of the most downloaded Japanese models of that year.
• Llama-3-ELYZA-JP — the 2024 series built on Llama 3, which the company reported scoring at the level of leading global models on Japanese-language benchmarks.
• Domain models — ELYZA has also published research on medical-specialised Japanese models and diffusion-based language models.

The approach keeps training costs far below those of frontier labs while producing models that handle honorifics, mixed scripts and Japanese business writing more naturally than general-purpose systems.

Inside KDDI

In March 2024 KDDI announced a capital and business alliance with ELYZA. From April that year KDDI held 43.4% of the shares and its subsidiary KDDI Digital Divergence Holdings held another 10%, making ELYZA a consolidated subsidiary. The deal was framed as a “swing-by IPO”: ELYZA uses KDDI’s customers, data centres and sales network to grow, while keeping a future stock listing as an option.

For KDDI the logic is straightforward. A carrier with tens of millions of mobile customers, a large enterprise IT business and a growing data-centre footprint needs its own AI capability rather than reselling only foreign models. ELYZA gives KDDI a team that can tune models for Japanese clients, including ones that cannot send data outside the country. Japanese press reported in 2026 that the two companies are also setting up a new venture in physical AI, the use of AI models to control robots and machines.

Products companies actually use

ELYZA’s commercial business centres on ELYZA Works, a no-code platform where staff build small AI apps for their own tasks, such as drafting replies, summarising meeting notes or checking contracts. In April 2026 the company added an AI agent feature that collects data from multiple Excel files, analyses it and produces documents automatically. This is the kind of spreadsheet-heavy reporting work that consumes a large share of office hours in Japanese firms.

Alongside the packaged product, ELYZA runs custom projects with large enterprises and public bodies, where it fine-tunes models on a client’s own documents and deploys them in a closed environment.

Why it matters for overseas partners

• Localisation partners — Global software vendors entering Japan can work with ELYZA-class models to make their products read naturally to Japanese users.
• Data residency — Banks, insurers and government-linked clients often require models that run in Japan. A KDDI-backed provider answers that requirement directly.
• Distribution — Through KDDI, ELYZA reaches a large base of corporate customers that would be slow to approach as a foreign startup.
• Investment watch — The swing-by IPO structure means ELYZA may eventually list, offering one of the few public ways to back a Japanese LLM specialist.

Outlook

ELYZA is not trying to beat the largest US labs on raw scale. Its bet is that Japanese companies will pay for models that understand their language and stay inside their compliance rules, delivered by a partner with a domestic telecom group behind it. With KDDI’s backing and a product that targets everyday office work, it has become one of the clearest examples of how Japan is building its own layer of the generative AI market.