Hugging Face · 2026-04-09 · major
Sentence Transformers v5.4 — multimodal embeddings and rerankers
The de-facto embedding library ships v5.4 with first-class multimodal support — encode and rerank text, images, audio and video through the same API, with Qwen3-VL, Llama Nemotron, BGE-VL and Jina rerankers working out of the box.
The most-used Python library for sentence embeddings adds native multimodal encoding and reranking across text, images, audio and video.
Key specs
| Version | v5.4.0 |
|---|---|
| License | Apache-2.0 |
| GitHub stars | 18.5k |
| Pretrained models on hf | 15,000+ |
What is it?
Sentence Transformers is the standard Python library for loading, running, and training embedding and reranker models. Version 5.4.0, released April 9 2026, adds native support for vision-language and other multimodal models so the same SentenceTransformer and CrossEncoder classes can encode text, images, audio and video, and score relevance between mixed-modality pairs. The Apache-2.0 repo is maintained by Hugging Face and hosts more than 15,000 pretrained models on the Hub.
How does it work?
The SentenceTransformer class was restructured to auto-detect input modality and route each item through the right encoder; a new Router module can compose separate encoders per modality in one pipeline. The CrossEncoder class was rewritten to inherit from a shared BaseModel with inspectable module chains and a new LogitScore module that unlocks generative rerankers. Text-only paths gained Flash Attention 2 input flattening, which concatenates variable-length sequences into flat tensors to cut padding waste when batch sequence lengths vary.
Why does it matter?
Multimodal RAG has been a DIY mess: every model shipped with its own loader, and mixing text and image embeddings meant stitching scripts together. With v5.4 you can drop Qwen3-VL-Embedding, NVIDIA Llama Nemotron, BGE-VL or Jina Reranker into an existing Sentence Transformers pipeline without changing calling code. In practice that means visual document retrieval and cross-modal search work the same way people already wrote their text-only retrieval code.
Who is it for?
Anyone building retrieval, RAG or search pipelines who wants images, audio, or video on equal footing with text.
Try it
pip install -U sentence-transformers