voyage-4-lite

voyage-4-lite is a lightweight text embedding model from the Voyage 4 series, engineered to balance retrieval quality with low latency and cost. Built with Matryoshka learning and quantization-aware training, it produces embeddings in 2048, 1024, 512, and 256 dimensions with options like int8, uint8, binary, and ubinary quantization. The model shares a unified embedding space with voyage-4-large and voyage-4, enabling cross-model compatibility where teams can index with a larger model for quality while querying with lite for speed.

Voyage AI designed lite to approach voyage-3.5 retrieval accuracy with fewer parameters, making it practical for cost-sensitive production pipelines. The shared embedding space across the Voyage 4 series means users aren't locked into a single model size—it supports flexible architectures like using voyage-4-large for document indexing while deploying lite for real-time queries. With 200M free tokens on signup and competitive per-token pricing, it's positioned for teams building retrieval-augmented systems at scale.