voyage-3.5

Voyage 3.5 is a text embedding model built for high-quality retrieval across a wide range of use cases. Its architecture centers on Matryoshka learning, which enables flexible output dimensions that teams can tune between 256 and 2048 depending on their storage and accuracy needs. The model also supports multiple quantization formats ranging from 32-bit floating point to compact binary precision, making it practical for serving at scale. With a 32,000-token context window and state-of-the-art standing in retrieval accuracy, it handles long documents and complex queries while remaining cost-efficient compared to alternatives.

The model was trained to excel in multilingual and domain-diverse retrieval scenarios, spanning 26 languages with robust performance in finance, legal, and code-heavy environments. Benchmarks show it outperforming OpenAI-v3-large by 8.26% on average across evaluated domains while maintaining significantly lower operational costs. This combination of accuracy and efficiency makes Voyage 3.5 particularly well-suited for large-scale retrieval pipelines, RAG systems, and applications where embedding quality directly impacts downstream results. The flexible quantization and dimension options further support teams optimizing for specific performance or resource constraints.