NII Releases Open-Source Japanese Language Model 'LLM-jp-4 33B' with 33 Billion Parameters

Summary

Japan's National Institute of Informatics (NII) has released LLM-jp-4 33B, a dense 33.2-billion-parameter Japanese large language model (LLM) under the Apache 2.0 license. The model is available in both base and SFT + DPO trained variants. NII claims that LLM-jp-4 33B outperforms its previous versions and OpenAI's gpt-oss-20b model across four different benchmarks.

Why it matters

This move strengthens Japan's 'sovereign AI' initiative. This locally developed, commercially viable model competes with Western open-weight models in Japanese and English tasks. The shift from the previous MoE-based 32B-A3B to a fully dense 33B architecture signals a deliberate investment in per-token quality over inference efficiency. This choice is critical for teams deploying the model on-premises. The Apache 2.0 license lowers the barrier for Japanese companies to develop applications on a non-US foundation model.

Highlights

• The dense 33B architecture replaces the previous MoE 32B-A3B; all parameters are activated for every token
• The Apache 2.0 license permits unlimited commercial use
• Outperforms gpt-oss-20b in MT-Bench (JP/EN) and AnswerCarefully safety evaluations

Critical take

All benchmark results are self-reported by NII using their own 'llm-jp-judge' framework. Furthermore, the headline comparison is made against a 20B-parameter open-weight model rather than a frontier system. Consequently, the 'superiority' claim remains limited in scope. The article does not explain the rationale for switching from MoE to a dense architecture. Therefore, it remains unclear whether the trade-off in inference cost is a deliberate choice or a constraint.