Embed v4.0

Embed v4.0 serves as a comprehensive solution for transforming diverse data types, including text, images, and interleaved combinations of both, into unified numerical vector representations. By design, it excels at capturing complex visual features from varied sources like PDFs, tables, and figures, which simplifies document processing by removing the need for traditional, rigid parsing methods. This architectural flexibility allows the model to function effectively across both English and multilingual environments, providing a robust foundation for applications requiring semantic search, clustering, and classification.

Built to handle large-scale information retrieval, the model supports a 128k context length and offers advanced compression techniques such as matryoshka embeddings, byte quantization, and binary quantization to optimize storage and performance. Its design lineage focuses on delivering state-of-the-art results across a wide range of academic and general-domain benchmarks, including the BEIR dataset. By enabling developers to fine-tune output dimensions and precision, the model provides a scalable path for integrating high-fidelity semantic understanding into modern data pipelines, ensuring it remains a practical choice for complex, multimodal information retrieval tasks.