Voyage Multimodal 3.5 is a next-generation multimodal embedding model built for retrieval over text, images, and videos. It embeds interleaved text and images (screenshots, PDFs, tables, figures, slides), and adds explicit support for video frames.
On-demand DeploymentDocs | On-demand deployments allow you to use Voyage Multimodal 3.5 on dedicated GPUs with Fireworks' high-performance serving stack with high reliability and no rate limits. |
voyage-multimodal-3.5 supports retrieval across text, images, screenshots, PDFs, figures, tables, slides, and videos. It is especially useful for datasets where documents combine text and visual information.
Videos are represented as ordered sequences of frames and embedded as visual inputs. This enables text-to-video retrieval, where a text query can search across scenes or video segments.
For longer videos, it is recommended to split videos into scenes, align segments with transcript timestamps when available, and reduce resolution or frame rate when needed to fit within the context limit.
voyage-multimodal-3.5 supports up to 32K tokens. For videos, every 1120 pixels counts as one token, up to the 32K token limit.
voyage-multimodal-3.5 supports 2048, 1024, 512, and 256 dimensional embeddings. These flexible dimensions are enabled by Matryoshka embeddings, allowing teams to balance retrieval accuracy, storage, and latency.
voyage-multimodal-3.5 supports multiple quantization options, including 32-bit floating point, signed and unsigned 8-bit integer, and binary precision. These formats can reduce storage and retrieval costs while minimizing accuracy loss.
