Voyage AI releases new rerankers and multimodal model on OpenRouter

Voyage AI has made rerank-3 and rerank-3-lite available on OpenRouter, positioning them as drop-in upgrades to rerank-2.5 and its lite counterpart. The practical upshot: developers can swap the model name in their API call without touching code or pipelines. Both models were designed from the ground up for long-document handling and code retrieval, two scenarios where earlier rerankers struggled to maintain accuracy as context length grew.

The bigger announcement is voyage-multimodal-3.5, a multimodal embedding model that can vectorize not just text, images and video separately, but also content that blends all three. That means a query containing a PDF screenshot, a table pulled from a slide deck and a short video clip returns a single vector representation preserving the relationships among them. According to the company, the model excels at mixed-modality searches of this kind, which are especially relevant for RAG over technical documents, financial reports or product documentation that includes diagrams and demo videos.

The multimodal model combines Matryoshka learning with quantization-aware training. Under the first approach, the vector learns to remain useful at lower dimensions, so it can be truncated from 2,048 dimensions to 1,024, 512 or 256 without retraining while still delivering reasonable performance. Quantization-aware training lets the weights be compressed to lower precision levels such as int8 or int4 with minimal quality loss. Together, the two techniques give developers control over the trade-off between storage cost, inference speed and quality without maintaining multiple separate model versions.

Availability on OpenRouter means all three models are accessible through a single API with one key, eliminating the need to manage separate accounts or hosting infrastructure. For teams already running RAG or semantic-search pipelines on OpenRouter, adoption amounts to zero friction — just change the `model` parameter in the request. The company also published a detailed technical blog post dated January 2026 explaining the architecture and benchmarks, though no independent comparison results against competing models such as BGE-M3 or Jina Embeddings v3 have been released at this stage.