Ruri v3 brings 8K-token embeddings to Japanese retrieval

Nagoya University’s CL Research Group has released Ruri v3, a family of Japanese sentence-embedding models built on ModernBERT-Ja. The models support inputs of up to 8,192 tokens, remove the MeCab preprocessing dependency used by many earlier Japanese encoders, and deliver strong reported results with as few as 30 million parameters.

Sentence-embedding models convert text into fixed-length vectors for semantic search, retrieval-augmented generation, clustering, classification, and duplicate detection. Ruri v3 targets Japanese workloads specifically, giving teams an Apache 2.0-licensed alternative to proprietary embedding APIs and larger open models.

An 8K-token rebuild

Ruri v3 updates the architecture and tokenizer in several practical ways:

• Longer inputs: The maximum sequence length rises from 512 tokens in Ruri v1 and v2 to 8,192 tokens.
• Larger vocabulary: A 100,000-token vocabulary replaces the earlier 32,000-token vocabulary, often reducing the number of tokens required for Japanese text.
• Faster attention: ModernBERT’s FlashAttention support can accelerate inference and fine-tuning on compatible hardware and software.
• Simpler tokenization: A SentencePiece-only tokenizer removes the need for an external Japanese morphological analyzer such as MeCab.

Removing morphological preprocessing simplifies container builds, cross-platform serving, and integration with standard sentence-transformers pipelines. The larger context window also accepts long articles, documentation pages, and conversations without truncating them at 512 tokens, although memory use and retrieval quality still depend on input length and batch size.