Voyage AI voyage-code-3

Acerca de Voyage AI

Fundado en 2023 · Palo Alto, California, USA

Voyage AI was founded in 2023 by Stanford CS professor Tengyu Ma and team, focused on best-in-class retrieval and reranking models for RAG, with a particular emphasis on domain-specific variants. The company has shipped specialised embeddings for finance (voyage-finance-2), law (voyage-law-2), multilingual (voyage-multilingual-2) and code (voyage-code-2 then voyage-code-3). voyage-code-3 launched in December 2024 as the successor to voyage-code-2 and quickly became the top-scoring code embedding model on the CoIR and CodeSearchNet benchmarks. In February 2025 Voyage AI was acquired by MongoDB for $220M, with voyage-code-3 now integrated into MongoDB Atlas Vector Search and recommended for code-aware AI agents and IDE-style retrieval workloads.

Arquitectura

Transformer bi-encoder specialised for code and technical text with Matryoshka heads

Voyage AI voyage-code-3 is a hosted embedding model specialised for source code, technical documentation, commit messages, issues, pull-request reviews and code-mixed natural language. It has the same 32,000-token context window as voyage-3 and supports Matryoshka-style heads at 256 / 512 / 1,024 / 2,048 dimensions, plus int8 / binary quantisation. Training used a contrastive retrieval objective on curated pairs covering more than 30 programming languages (Python, JavaScript, TypeScript, Go, Rust, Java, C/C++, C#, SQL, Bash, etc.) together with technical natural-language text, with deliberate emphasis on code-to-text and text-to-code retrieval as well as code-clone detection. Voyage reports voyage-code-3 outperforming OpenAI text-embedding-3-large by 13.8 points on average across CoIR sub-tasks while costing the same. The model is offered through the Voyage API and natively in MongoDB Atlas Vector Search after the February 2025 acquisition.

Parámetros

Undisclosed

Contexto

32,000 tokens

Capacidades

• Top-tier code embedding model on CoIR and CodeSearchNet
• 30+ programming languages including Python, JavaScript, TypeScript, Go, Rust, Java
• Text-to-code and code-to-text retrieval (e.g. find function from docstring)
• 32,000-token context window for full-file embedding
• Matryoshka heads at 256 / 512 / 1024 / 2048 dimensions
• int8 / binary quantisation for cheap storage
• MongoDB Atlas Vector Search integration
• Best for: code-aware AI agents, IDE retrieval, repository search, ticket-to-code linking

Entrenamiento y licencia

Not disclosed. Voyage describes 'curated pairs covering 30+ programming languages and technical natural-language text' with contrastive negatives.

Licencia: Proprietary commercial API. Available standalone and bundled with MongoDB Atlas Vector Search.

Pruebas de seguridad: Embeddings are not subject to content-filter constraints. Voyage publishes general fairness statements.