LLM-jp-3

LLM-jp-3 is the third-generation open Japanese large language model series from the Research and Development Center for Large Language Models at Japan's National Institute of Informatics. Trained on the llm-jp-corpus v3, the series spans eight sizes from 150M to 172B parameters, with the 172B flagship trained on about 2.1 trillion tokens and reported to surpass GPT-3.5 on Japanese evaluations.

True to the project's open-science mandate, training data, code, and evaluation frameworks are published. Models at 13B and below are Apache 2.0; the 172B weights require access approval with redistribution restrictions, so check terms before building on the flagship. Instruct2 (SFT) and Instruct3 (DPO) alignment generations are available across sizes.