Codestral Embed

Codestral Embed, launched May 28, 2025, is Mistral AI's first embedding model — and it is code-specific. Available via the Mistral API as codestral-embed-2505 at $0.15 per million tokens, it supports configurable output dimensions (up to 3072, Matryoshka-ordered so you can truncate cleanly) and precisions, letting teams trade retrieval quality against storage cost.

Mistral reported it significantly outperforming Voyage Code 3, Cohere Embed v4.0, and OpenAI's Text Embedding 3 Large on SWE-Bench and Text2Code retrieval, and claimed that even at 256 dimensions with int8 precision it beats competitors' full-size outputs. Use cases span RAG over codebases, semantic code search, duplicate detection, and repository analytics.

Honest caveats: the benchmark claims are Mistral's own, there are no open weights (API-only, though on-premise deployments are negotiable), and it is code-tuned — reach for mistral-embed or a general model for prose. For code RAG on the Mistral platform, it is the obvious pairing with Codestral generation models.