Granite Embedding 97M Multilingual R2
Granite Embedding 97M Multilingual R2 is a ModernBERT-based bi-encoder for retrieval. It produces 384-dimensional vectors. Its language pretraining covers more than 200 languages, with retrieval training focused on over 52 languages and code.
Model design and output
The compact model prunes a 22-layer design to 12 layers and uses a 180,000-token vocabulary. Distillation transfers retrieval behavior from a larger model. Query and document vectors are compared for relevance; the model does not generate answers.Granite Embedding 97M Multilingual R2 model card
Practical use
Practical guidance. Use this model when vector storage and embedding throughput are important. Evaluate cross-language queries against the actual languages in the collection. A multilingual label does not guarantee equal retrieval quality for every language, script or technical domain.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ibm-granite/granite-embedding-97m-multilingual-r2 |
| Context window | 32,768 tokens Published maximum input length. Granite Embedding 97M Multilingual R2 model card |
| License | Apache 2.0Granite Embedding 97M Multilingual R2 model card |
| Parameters | 97MGranite Embedding 97M Multilingual R2 model card |
| Vector size | 384 dimensionsGranite Embedding 97M Multilingual R2 model card |
Related models
- Granite Embedding 311M Multilingual R2
IBM ยท Granite Embedding
A multilingual ModernBERT retrieval encoder with 768-dimensional output and smaller Matryoshka vector options.
Launch date not confirmedDownloadable