AI provider

LLM-jp / NII LLMC

LLM-jp is a Japanese research community and model-development project led by the National Institute of Informatics. Its Research and Development Center for Large Language Models, LLMC, develops models and shares research methods, data and tools.

Explore models

Company & leadership

Project established: May 2023, when NII started the LLM-jp research group.
Research center established: 1 April 2024. LLMC is part of the National Institute of Informatics, within the Research Organization of Information and Systems (ROIS).
Institutional leadership: Sadao Kurohashi is LLMC director and NII director-general. Akiko Aizawa and Koichi Takeda are vice directors. The member page is dated 1 November 2026.
Founding organization: NII established the research group and center. LLM-jp is a collaborative research project with university and company participants, rather than a separately incorporated company.
Office: Palace Side Building, 3F, 1-1-1 Hitotsubashi, Chiyoda-ku, Tokyo 100-0003, Japan.
Public research support: LLMC was established to implement MEXT's research hub program for transparency and reliability in generative AI.

The study group became a research center

NII started LLM-jp in May 2023. Its first 13B model was released on 20 October 2023 after training began on the mdx computing platform in July. The project included researchers from academia and industry before LLMC was established in April 2024.

The project publishes methods and failed experiments

LLMC's stated approach includes release of models, development data, tools and technical discussions. It also publishes failures. The 172B alpha releases are a concrete example: their announcement identifies an AdamW optimizer setting that prevented the expected performance.

Research now covers reasoning, images and conversation

The September 2026 LLM-jp-4.1 release added new post-training for three reasoning models. The official release index also includes vision-language models and Japanese full-duplex speech. These are research outputs with model-specific access terms, not a single managed API catalog.

Model releases and research variants

Practical guidance. The entries below separate pretrained models from instruction and preference-tuned releases. File formats stay with their model. The Drop-Upcycling and Optimal Sparsity articles retain the named experimental variants. Tokenizer-only repositories and untrained initialization files do not receive model pages.