LLM-jp-4 32B-A3B

LLM-jp-4 32B-A3B is the mixture-of-experts flagship of the LLM-jp-4 series, released in 2026 by the NII-organized LLM-jp consortium. With 32 billion total parameters and 3 billion active per token, it targets efficient high-quality Japanese generation for research and enterprise use.

Like the 8B sibling, it ships under Apache 2.0 with base, instruct, and thinking variants, and coverage places it ahead of GPT-3.5-turbo on Japanese benchmarks. The thinking variant scores 7.57-7.82 on Japanese MT-Bench in vendor-published evaluations. It represents the largest open Japanese model in the LLM-jp-4 family.