Voyage Code 3

voyage-code-3, released December 4, 2024, is Voyage AI's code-tuned embedding model — and one of the most widely deployed retrieval models in the AI coding ecosystem, serving as the retrieval engine behind assistants like Cursor, Continue, and Cody. It offers a 32K-token context window with Matryoshka-configurable output dimensions (256/512/1024/2048) and multiple precisions.

Voyage reported it beating OpenAI's text-embedding-3-large by nearly 14% on average across code retrieval datasets, with particular strength on CodeSearchNet. The flexible dimensions let teams store embeddings at a third the size of competitors at comparable accuracy — a practical win for large codebases.

Honest caveats: it is API-only (no open weights), benchmark claims are Voyage's, and the 2026 voyage-code-4 release has since superseded it for agent-heavy workloads. Pricing and rate limits are tiered by account. For code RAG where retrieval quality is the bottleneck, it remains a proven default.