Hugging Face has introduced multimodal support in Sentence Transformers, allowing developers to embed and rerank content that mixes text and images. The update extends the library's familiar interface so existing retrieval workflows can incorporate visual data without switching to a separate framework.
According to the blog post, these models are designed for pipelines where queries and documents may live in different modalities. Reranker models can take a query and candidate pairs across text and images, then return relevance scores to improve ranking after an initial recall stage.
Because there is only one source, there are no conflicting viewpoints to note. The post frames the change as a natural step forward for the library, bringing multimodal search closer to the same developer experience as text-only retrieval.