The archive · AI & Models · Product decision · 2018–2024
ELYZA bets Japanese-language LLMs can beat English-first models; KDDI takes control
Tokyo's ELYZA built Japanese-optimized LLMs claiming GPT-4-level Japanese; KDDI made it a subsidiary in 2024 to sell enterprise GenAI
ELYZA Inc. [株式会社ELYZA]
What the business is
Tokyo-based LLM startup (University of Tokyo Matsuo lab spin-off) that builds Japanese-language large language models and enterprise GenAI services (ELYZA LLM for JP, ELYZA Works).
Starting capital：KDDI invested 'several billion yen' [数十億円, on the lower end of the range per KDDI's CEO] for the 53.4% stake; exact amount undisclosed.
How it started
ELYZA was founded in September 2018 by members of the University of Tokyo Matsuo lab and focused on LLM research and enterprise deployment from summer 2019. On 2024-03-12 it released the 70B-parameter ELYZA-japanese-Llama-2-70b, claiming the highest Japanese-language accuracy among domestic models, matching global ones.
What happened
On 2024-03-18 ELYZA, KDDI and KDDI Digital Divergence signed a capital/business alliance: from 2024-04-01 KDDI held 43.4% and KDDI Digital Divergence 10.0%, making ELYZA a consolidated subsidiary while keeping a swing-by IPO path. KDDI said it would build roughly ¥100 billion of compute infrastructure for it and roll out enterprise/local-government GenAI services from spring 2024.
How it ended up
Still running as KDDI's enterprise-GenAI engine: it supplies Japanese LLMs and AI services to enterprises and local governments, with a swing-by IPO as the stated goal.
Background
ELYZA was founded in September 2018 by members of the University of Tokyo Matsuo lab, and from summer 2019 it focused on LLM research and enterprise deployment. Its bet was that Japanese enterprises need models truly fluent in Japanese — something English-first frontier models left on the table — so a small Tokyo lab could win by optimizing open models for Japanese instead of chasing the global frontier.
The approach skipped the most expensive step: instead of pre-training a model from scratch, ELYZA took Meta's open Llama models and added large-scale Japanese pre-training and instruction tuning, with exclusive access to 13% of Japan's AIST ABCI supercomputer supported by METI. In March 2024 it released the 70B ELYZA-japanese-Llama-2-70b, claiming the top Japanese-language performance among domestic models.
On 2024-03-18 ELYZA, KDDI and KDDI Digital Divergence signed a capital/business alliance: from April 1, 2024, KDDI held 43.4% and KDDI Digital Divergence 10.0% of ELYZA, making it a consolidated subsidiary while keeping a swing-by IPO path. KDDI — which had no frontier model of its own while NTT (Tsuzumi) and SoftBank built theirs — said it would invest roughly ¥100 billion in compute infrastructure and sell enterprise and local-government GenAI services from spring 2024.
On 2024-06-26 ELYZA released Llama-3-ELYZA-JP-70B and Llama-3-ELYZA-JP-8B, claiming the 70B model matched or exceeded GPT-4, Claude 3 Sonnet and Gemini 1.5 Flash on two Japanese benchmarks (ELYZA Tasks 100, Japanese MT-Bench) as of end-June 2024, and open-sourced the 8B model on Hugging Face.
What has to be true
- Japanese is hard for English-first models, so generic global LLMs left real accuracy on the table for Japanese enterprises.
- Building on open Llama models skipped the billion-dollar pre-training step while still producing differentiated Japanese models.
- Matsuo-lab pedigree plus METI/AIST supercomputer access gave credibility and compute a startup could not buy.
- KDDI lacked an in-house frontier model while NTT and SoftBank built their own, so buying ELYZA gave instant domestic-LLM credibility.
- The swing-by IPO structure let ELYZA keep founder-led growth inside a carrier's compute and sales assets.
What can be applied
A small lab can compete by owning a niche giants neglect — Japanese fluency — and by standing on open models to skip the costliest step; the exit was a carrier needing domestic AI credibility.
Aftermath
As of late June 2024, ELYZA operates as KDDI's consolidated subsidiary: it released Llama-3-ELYZA-JP-70B/8B on June 26, 2024 (the 70B offered via API and demo site, the 8B open-sourced on Hugging Face), and is rolling out enterprise and local-government GenAI services with KDDI, including contact-center LLMs developed with Altius Link; a swing-by IPO remains the stated goal.
Sources
- ELYZAとKDDIグループ、生成AIの社会実装に向け資本業務提携を締結
- KDDI、生成AIのELYZA子会社化 最高性能の日本語LLMを展開
- 「GPT-4」超えを達成した国産日本語LLM、AIスタートアップのELYZAが開発・公開
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card