Mistral: Codestral Embed 2505

Mistral Codestral Embed 2505 is an embeddings model from Mistral, released on 30 October 2025 and published on api14 under the model id mistralai/codestral-embed-2505. It sits in the embeddings category rather than the chat category, which means it is not built to hold a conversation or generate written answers. Instead it turns input into vectors that represent meaning, and Mistral describes it as specially designed for code.

The model accepts text as its only input and returns embeddings as its only output, with a context window of 8192 tokens. That is enough room to embed a substantial function, a class, or a documentation section in a single pass. It declares no supported sampling parameters, which is typical for embedding endpoints where temperature and similar controls do not apply; you send your input and read the resulting vector. The max output field is listed as 6553.

Because it is tuned for code, the most direct use cases are the ones Mistral names: embedding code databases and repositories, and powering coding assistants that need strong retrieval. In practice that covers semantic search across a monorepo, deduplication of near-identical snippets, clustering files by similarity, and building the retrieval layer that feeds relevant code into a coding assistant. General text embedding is possible since input is plain text, but the model's stated focus is source code and the retrieval workloads around it.

Calling it through api14 works the same way as any other model on the platform, because api14 exposes an OpenAI-compatible API. You point your existing OpenAI client at the api14 base URL, supply your api14 key, and pass mistralai/codestral-embed-2505 as the model id on the embeddings endpoint. No extra parameters are listed as supported for this model, so a straightforward request containing the model id and your input text is what you need. If you already have embedding code written against the OpenAI SDK, the change is essentially the base URL, the key, and the model string.

The main thing to weigh is the division of labour. This model produces vectors, not prose, so a task that needs generated explanations, code completion, or agent behaviour requires a chat model alongside it, with Codestral Embed handling the retrieval step. Where your workload is purely semantic matching over code or text and fits within the 8192 token context, this model is the appropriate choice.