OpenSearch open-sources its first dense embedding model and an asymmetric sparse retriever

tomaarsen · x · 2026-11-13

The OpenSearch project released two Apache 2.0 models: an asymmetric sparse encoding model with a smaller query tower than document tower (opensearch-neural-sparse-encoding-asym-v1), plus the project's first dense embedding model. Both are Sentence Transformers-native with strong reported results for their sizes. tomaarsen demos sparse retrieval in a few lines: encodequery/encodedocument, direct similarity, and token-level weight decoding (solar 2.788, panel 2.431...) for interpretability. Sparse retrieval remains underrated for explainability and out-of-domain generalization.

Original post →

OpenSearch open-sources its first dense embedding model and an asymmetric sparse retriever

More from Models