LLM-jp-4

LLM-jp-4 is the fourth-generation open-source Japanese LLM from LLM-jp, the 1,000-plus-member research consortium coordinated by Japan's National Institute of Informatics (NII). The family includes a dense 8B model and a 32B-A3B mixture-of-experts model (32B total, ~3.8B active, 128 routed experts), with base, instruct, and thinking variants, all under Apache 2.0 on Hugging Face. A follow-up, LLM-jp-4.1, added a 33B dense variant and day-one GGUF quants in September 2026.

The models were trained on a large corpus of Japanese text including legal, medical, and governmental sources, and score strongly on Japanese benchmarks like JGLUE. The 4.1 release drew attention for shipping official quantized GGUF files (the 8B Q4_K_M fits in 5.5GB) on launch day, unusual among research releases. Release date is era-level 2026-05, based on earliest Hugging Face presence.