codestral-embed

Overview. Codestral Embed is Mistral AI's first code-specialised embedding model, released on 28 May 2025 and distributed only through Mistral's API and Batch API as a closed-weight product. It ingests text or code up to an 8,192-token window and returns a numeric embedding vector rather than generated text. Developers reach for it when they need code representations for search, clustering, or retrieval over a codebase, where a general-purpose embedding model would capture semantics less accurately.

Key strengths. The model emits matryoshka-style vectors: the default 1,536 dimensions can be truncated to any smaller count while preserving the most relevant components first, up to a maximum of 3,072 dimensions. Output can be requested as float, int8, uint8, binary, or ubinary, which lets a developer trade storage and comparison cost against precision without re-embedding. Mistral positions it for code-retrieval workloads and recommends chunking source at roughly 3,000 characters with 1,000-character overlap to keep retrieval quality stable across file boundaries.

When to use it. Use this for code retrieval and semantic search over repositories, where a code-specialised embedding produces more useful nearest-neighbour results than a text model. The matryoshka output makes it a good fit when you need to vary embedding dimension per index or tier — full-width vectors for a high-precision store, truncated binary vectors for a fast first-pass filter. The 8,192-token window suits whole functions or multi-file chunks, and the code-specialised objective is a better match for identifier- and syntax-heavy text than a general embedding endpoint.

Limitations. The model is embedding-only: it returns vectors, so it cannot generate text, call tools, or serve as a chat or reasoning model despite carrying the Codestral name. Input is text and code only — there is no image or audio support. The input window is 8,192 tokens, and Mistral has not published a parameter count, architecture, or training cutoff, so the model's internal capacity and recency cannot be assessed from documentation.

Sovereignty note. Codestral Embed is built by Mistral AI, a French creator with no CLOUD Act exposure, and is served only through Mistral's API with no public weights. An EU-resident endpoint tier is available for Sovereign routing where data must remain within the EU.