LLM-jp-4 33B

LLM-jp-4 is the latest generation of Japan's LLM-jp project, a national research consortium led by the National Institute of Informatics (NII), continuing the Apache 2.0-licensed open-weights line that began with LLM-jp-1. The 33B dense model targets high-quality Japanese text generation, with a thinking/reasoning variant (llm-jp-4-33b-thinking) for harder problems, and quantized distributions (e.g. GPTQ INT4 at ~20.6GB) make it runnable on 48GB-class GPUs. Like its predecessors it is built as public research infrastructure for Japanese NLP rather than a commercial product.

Limitations: Documentation is primarily Japanese with sparse English materials, raising the bar for non-Japanese-speaking users. As an academic consortium release it lacks the polished chat fine-tuning and safety alignment of commercial labs' models. The 33B dense size demands serious GPU memory even quantized, and independent benchmark coverage in English-language media is thin.

Limited public benchmarks; positioned as a strong Japanese writer