On October 2, ELYZA, an AI development company within the KDDI Group, released the AI models "ELYZA-Thinking-1.0-llm-jp-4-33b" and "ELYZA-Thinking-1.0-llm-jp-4-32b-a3b" for free. These models are based on the "LLM-jp-4" series, an open foundation model developed by the National Institute of Informatics (NII), and have been enhanced through post-training.
Model ReleasesELYZAELYZA-Thinking-1.0-llm-jp-4-33bELYZA-Thinking-1.0-llm-jp-4-32b-a3b
ELYZA Releases Two New AI Models Claiming "Fully Domestic Foundation," Enhancing Performance Based on LLM-jp-4
This article is a translation. Read the Japanese original
ELYZA describes these models as a "fully domestic foundation" and states that they conducted intermediate and post-training focusing primarily on improving Japanese language processing performance. They mentioned that in some benchmarks, the models achieved scores exceeding the "LLM-jp-4.1" series—the post-training improved version of LLM-jp-4 announced by NII on September 28.
Both of the released models are licensed under Apache 2.0, allowing for commercial use. They have been available on Hugging Face since October 2.
This announcement is part of the initiatives of "ELYZA RSI Research," a research organization established by ELYZA on the same day. The organization aims to promote "Recursive Self-Improvement (RSI)," where AI iteratively improves its own capabilities and development processes. It should be noted that as of the time of the announcement, it is in the elemental research stage preceding RSI and has not achieved RSI itself.
Sources
- 「完全国産」AI、KDDI傘下のELYZAが無料公開 「LLM-jp-4」ベースに性能強化 （ITmedia AI+、2026-10-02）
- プレスリリース