Voyage Multimodal-3

Voyage AI's multimodal embedding models transform unstructured data from multiple modalities into a shared vector space. Unlike traditional models like CLIP, which process text and images separately, Voyage's models use a single backbone to directly vectorize inputs containing interleaved text and images. This eliminates the need for complex text extraction or ETL pipelines and removes modality bias in mixed-modality searches. The models support rich visual data such as PDF screenshots, slide decks, tables, figures, photos, and videos (video support via voyage-multimodal-3.5). They offer a context length of 32,000 tokens and various embedding dimensions (default 1024, with options 256, 512, 2048). Voyage multimodal embeddings are accessible through a Python API via the voyageai package, with usage-based pricing and a generous free tier (first 200M text tokens and 150B pixels free).