Codestral Embed

Mistral AI has launched Codestral Embed, its first embedding model specialized in code data retrieval. This model outperforms Voyage Code 3, Cohere Embed v4.0, and OpenAI's large embedding models.

Users can choose various dimensions and precisions to balance retrieval quality and storage costs. In particular, the combination of 256 dimensions and int8 precision alone outperforms competitor models. The embedding dimensions are sorted by relevance, allowing users to select as many leading dimensions as needed to flexibly adjust cost and quality.

Key use cases include the following:

• RAG (Retrieval-augmented generation): Efficient context retrieval for code completion, editing, and explanation tasks.
• Semantic code retrieval: Precise code snippet retrieval via natural language or code queries.
• Similarity search and duplicate detection: Identifying functionally similar code and detecting copy-pasting for license compliance.
• Semantic clustering and code analysis: Unsupervised grouping of code by function or structure.

Codestral Embed is available via API (codestral-embed-2505) at $0.15 per 1 million tokens, with a 50% discount applied when using the Batch API. For optimal retrieval performance, using 3,000-character chunks with a 1,000-character overlap is recommended.