Codestral Embed
2025-05-28 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI launches Codestral Embed, its first embedding model specialized in code, announced as outperforming Voyage Code 3, Cohere Embed v4.0 and OpenAI's large embedding model for code retrieval.
What changes
- Mistral's first embedding model dedicated to code, designed for retrieval on real repositories.
- Configurable dimensions and precisions (e.g. 256 dimensions in int8); dimensions are ordered by relevance, so you can truncate to the first n.
- Even at 256 dimensions int8, announced as outperforming competing models.
- Use cases: RAG for coding assistants, semantic search, duplicate detection, clustering.
### Specifications - Context: 8192 tokens; recommended chunking into blocks of 3000 characters with 1000 overlap. - License: proprietary (Premier model); on-premises deployment on request. - Availability: La Plateforme API and batch API. - API identifier: `codestral-embed-2505`. ### Pricing | Model | Price (USD / M tokens) | |---|---| | Codestral Embed | 0.15 | | Batch API | 50% discount | ### Benchmarks The charts in the post are images; no numerical score in text. Categories evaluated: SWE-Bench Lite (file retrieval), Text2Code (GitHub), Text2SQL, algorithms, data science. ### Key takeaways A specialized embedding for indexing large codebases, with fine tuning of the trade-off between quality and storage cost.
Links
Tags : LLM, Mistral, Embedding, Code, RAG