Granite Embedding 30M Sparse
Granite Embedding 30M Sparse is an English retrieval encoder in IBM's experimental Granite releases. It produces weighted token expansions instead of a fixed-size dense vector.
Research release
Model design and output
The model uses a sparse bi-encoder approach and supports SPLADE-style retrieval integration. Its output can be indexed using sparse search systems. It differs from the dense Granite embedding releases in both representation and retrieval infrastructure.Granite Embedding 30M Sparse model card
Practical use
Practical guidance. Compare sparse retrieval with your lexical and dense search baselines. Inspect which expanded terms cause useful matches and which introduce noise. Sparse and dense scores use different scales, so combining them requires an explicit ranking method.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | ibm-granite/granite-embedding-30m-sparse |
| License | Apache 2.0Granite Embedding 30M Sparse model card |
| Parameters | 30MGranite Embedding 30M Sparse model card |
| Retrieval format | Sparse token-weight dictionaryGranite Embedding 30M Sparse model card |
Related models
- Granite Embedding Small English R2
IBM ยท Granite Embedding
A 47M ModernBERT English retrieval encoder that produces 384-dimensional vectors.
Launch date Downloadable