Fireworks AI announced Voyage Multimodal 3.5, a retrieval‑focused embedding model that handles interleaved text and visual content. It works with screenshots, PDFs, tables, figures, slides and video frames. The model supports up to 32 000 tokens, counting each 1120 pixels of video as one token. Users can choose embedding dimensions of 256 to 2048, thanks to Matryoshka embeddings, and select from several quantization formats to cut storage costs.
Why it matters
Teams that need to search mixed‑media datasets can use Voyage Multimodal 3.5 to find relevant text or video clips without building separate indexes.
