14

Sentence Embeddings, Retrieval, and Reranking

In the Hugging Face ecosystem, the fields of Japanese sentence embeddings, retrieval, and reranking are developing rapidly. With the spread of RAG (Retrieval Augmented Generation), demand for high-quality Japanese text embedding models is growing.

14.1 Sentence Embedding Models (Sentence Similarity)

In the sentence-similarity category on Hugging Face, many models used for semantic similarity calculation of Japanese text are published.

Model HF DL Likes Description
ruri-v3-310m (cl-nagoya) 634k 89 ModernBERT-Ja based; up to 8192 tokens
plamo-embedding-1b (pfnet) 486k 49 1B-parameter embeddings
ruri-v3-30m (cl-nagoya) 126k 12 Lightest Ruri
ruri-v3-130m (cl-nagoya) 93k 9 Mid-size Ruri v3
GLuCoSE-base-ja (pkshatech) 31k 34 LUKE-based sentence embeddings
ruri-v3-70m (cl-nagoya) 14k 5 Small Ruri v3
ruri-large (cl-nagoya) 8k 47 Large-scale Ruri
GLuCoSE-base-ja-v2 (pkshatech) 6k 24 CPU-friendly version
ruri-base (cl-nagoya) 5k 13 General-purpose Japanese embeddings
JaColBERTv2 (bclavie) 2k 17 ColBERT-based Japanese retrieval

The Ruri series (cl-nagoya / Nagoya University): The leading family of general-purpose Japanese text embedding models with the highest usage in the sentence-similarity category. In v3, it supports input of up to 8,192 tokens based on ModernBERT-Ja, with multiple sizes from 30M to 310M. It achieves scores of 74.5-77.2 on the JMTEB benchmark. ruri-v3-310m (634k/week) continues at the top, with ruri-v3-30m (126k/week) also widely used, making the series the standard choice for Japanese embedding models. The mid-size ruri-v3-130m (93k/week) and ruri-v3-70m (14k/week) have also taken hold, and this ability to pick a size along the accuracy/inference-cost trade-off is a key strength of the Ruri series.

PLaMo-Embedding-1B (Preferred Networks): A large-scale embedding model with 1B parameters that shows high performance on the JMTEB benchmark. It is available for free use under the Apache v2.0 license.

sarashina-embedding (SB Intuitions): A 1.2B-parameter embedding model based on Sarashina2.1, trained with multi-stage contrastive learning, and achieves state-of-the-art scores on JMTEB.

From an architectural perspective, the Ruri series uses a Siamese/bi-encoder architecture suited for real-time search and caching. For model size selection: use ruri-v3-30m (30M) for low-latency/edge deployment; ruri-base or ruri-v3-310m (310M, up to 8,192 tokens) for a balance of quality and speed; and sarashina-embedding-1b for maximum accuracy (state-of-the-art on JMTEB).

ruri-v3-310m (cl-nagoya)
ModernBERT-Ja based; up to 8192 tokens
634k 89
plamo-embedding-1b (pfnet)
1B-parameter embeddings
486k 49
ruri-v3-30m (cl-nagoya)
Lightest Ruri
126k 12
ruri-v3-130m (cl-nagoya)
Mid-size Ruri v3
93k 9
GLuCoSE-base-ja (pkshatech)
LUKE-based sentence embeddings
31k 34
ruri-v3-70m (cl-nagoya)
Small Ruri v3
14k 5
ruri-large (cl-nagoya)
Large-scale Ruri
8k 47
GLuCoSE-base-ja-v2 (pkshatech)
CPU-friendly version
6k 24
ruri-base (cl-nagoya)
General-purpose Japanese embeddings
5k 13
JaColBERTv2 (bclavie)
ColBERT-based Japanese retrieval
2k 17
Ruri series
The Ruri series (cl-nagoya / Nagoya University): The leading family of general-purpose Japanese text embedding models with the highest usage in the sentence-similarity category. In v3, it supports input of up to 8,192 tokens based on ModernBERT-Ja, with multiple sizes from 30M to 310M. It achieves scores of 74.5-77.2 on the JMTEB benchmark. ruri-v3-310m (634k/week) continues at the top, with ruri-v3-30m (126k/week) also widely used, making the series the standard choice for Japanese embedding models. The mid-size ruri-v3-130m (93k/week) and ruri-v3-70m (14k/week) have also taken hold, and this ability to pick a size along the accuracy/inference-cost trade-off is a key strength of the Ruri series.
PLaMo-Embedding-1B
PLaMo-Embedding-1B (Preferred Networks): A large-scale embedding model with 1B parameters that shows high performance on the JMTEB benchmark. It is available for free use under the Apache v2.0 license.
sarashina-embedding
sarashina-embedding (SB Intuitions): A 1.2B-parameter embedding model based on Sarashina2.1, trained with multi-stage contrastive learning, and achieves state-of-the-art scores on JMTEB.

14.2 Feature Extraction

Model HF DL Likes Description
sentence-bert-base-ja-mean-tokens-v2 (sonoisa) 53k 51 Japanese Sentence-BERT v2
japanese-clip-vit-b-16 (rinna) 40k 25 Japanese CLIP
clip-japanese-base (LINE) 34k 30 Japanese CLIP (trained on 1B pairs)
sentence-bert-base-ja-mean-tokens (sonoisa) 14k 11 Japanese Sentence-BERT v1
clip-japanese-base-v2 (LINE) 2k 18 Improved version of clip-japanese-base
sarashina-embedding-v2-1b (sbintuitions) 2k 28 JMTEB state-of-the-art, 1792 dimensions
t5-base-japanese (sonoisa) 1k 57 Trained on 100GB Wikipedia/OSCAR

Multiple Japanese CLIP models are also published, with rinna/japanese-clip-vit-b-16 (40k/week DL) and LINE/clip-japanese-base (34k/week DL) being representative. They are used for learning the correspondence between images and text. The successor clip-japanese-base-v2 (2k/week DL) is also available.

sentence-bert-base-ja-mean-tokens-v2 (sonoisa)
Japanese Sentence-BERT v2
53k 51
japanese-clip-vit-b-16 (rinna)
Japanese CLIP
40k 25
clip-japanese-base (LINE)
Japanese CLIP (trained on 1B pairs)
34k 30
sentence-bert-base-ja-mean-tokens (sonoisa)
Japanese Sentence-BERT v1
14k 11
clip-japanese-base-v2 (LINE)
Improved version of clip-japanese-base
2k 18
sarashina-embedding-v2-1b (sbintuitions)
JMTEB state-of-the-art, 1792 dimensions
2k 28
t5-base-japanese (sonoisa)
Trained on 100GB Wikipedia/OSCAR
1k 57

14.3 Text Reranking

Japanese models for reranking (re-ordering search results) in RAG pipelines are also well-represented.

Model HF DL Likes Description
japanese-reranker-cross-encoder-small-v1 (hotchpotch) 156k 5 CrossEncoder; rank 11 across all HF Japanese models
ruri-v3-reranker-310m (cl-nagoya) 137k 18 ModernBERT-Ja based
japanese-reranker-xsmall-v2 (hotchpotch) 28k 7 v2 version; ONNX support
japanese-reranker-cross-encoder-xsmall-v1 (hotchpotch) 16k 7 Multiple sizes from xsmall to large
ruri-reranker-small (cl-nagoya) 3k 2 Small reranker from the Ruri v1 line

The Japanese reranker series by hotchpotch provides multiple sizes from xsmall to large, and has been evaluated on JQaRA, JaCWIR, MIRACL, and JSQuAD benchmarks. In this round, japanese-reranker-cross-encoder-small-v1 (156k weekly downloads) moved back ahead of ruri-v3-reranker-310m (137k), retaking the top spot in the text-ranking category. They rank #11 and #13 respectively across all Japanese NLP models on Hugging Face, indicating the scale of demand for reranking in RAG systems. japanese-reranker-xsmall-v2 (28k) and japanese-reranker-cross-encoder-xsmall-v1 (16k) follow. In v2, support for Flash Attention 2 and ONNX/quantized formats is also provided.


japanese-reranker-cross-encoder-small-v1 (hotchpotch)
CrossEncoder; rank 11 across all HF Japanese models
156k 5
ruri-v3-reranker-310m (cl-nagoya)
ModernBERT-Ja based
137k 18
japanese-reranker-xsmall-v2 (hotchpotch)
v2 version; ONNX support
28k 7
japanese-reranker-cross-encoder-xsmall-v1 (hotchpotch)
Multiple sizes from xsmall to large
16k 7
ruri-reranker-small (cl-nagoya)
Small reranker from the Ruri v1 line
3k 2