About
LLM-jp is a collaborative Japanese project, led by the National Institute of Informatics, that builds large language models with the entire development process made public — data selection, training code, and weights — so results can be verified and reproduced. The latest LLM-jp-4 family includes an 8B dense model and a 32B-A3B mixture-of-experts model, each trained from scratch on roughly 12 trillion tokens with a strong Japanese corpus. Earlier generations (1.3B to 172B) are also published, making it a foundation for sovereign Japanese AI and LLM research.
Download the weights from the llm-jp organization on Hugging Face (for example the LLM-jp-4 collection linked above) and run them locally with Hugging Face Transformers or a compatible inference stack — no account or API key needed for the public models.
More in AI Models
View allShisa AI
Open-weight bilingual Japanese–English models that punch above their size in Japanese.
Sakana Fugu
One API that orchestrates a pool of frontier LLMs as a multi-agent system.
ChatGPT
The most versatile AI assistant, with image generation built in.