voyage-code-3
Voyage Code 3 is a code retrieval model trained with text, code and mathematical material. It supports variable output dimensions.
Model design
The release describes contrastive pairs such as docstrings with code and code with related code. Its training mix spans more than 300 programming languages. The proportion of text and code is a deliberate part of training, rather than treating all repository text alike.voyage-code-3 code embeddings
Practical use
Practical guidance. Store repository, path and revision with each embedded passage. Related code from an old revision may fail to explain the current program. Evaluate natural-language requests separately from code-to-code lookup because their failure modes can differ.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API |
| Context window | 32,000 tokens Per input text. voyage-code-3 code embeddings |
| Atlas lifecycle | Legacy since 13 August 2026MongoDB Voyage AI model lifecycle: Atlas scope and release history |
| Output dimensions | 1,024 default; 256, 512 or 2,048 optionalMongoDB Voyage AI model specifications |
Related models
- voyage-code-4
Voyage AI · Voyage Code
Code embeddings trained with completed pull requests and repository tasks.
Launch date API - voyage-code-2
Voyage AI · Voyage Code
Earlier code retrieval encoder for natural-language and code queries.
Launch date API