![Mistral]()
Codestral has launched Codestral Embed, its first embedding model specifically optimized for code. Designed to excel in real-world code retrieval use cases, Codestral Embed outperforms current market leaders and sets a new standard in code embedding technology.
Superior Performance Compared to Competitors
Codestral Embed significantly surpasses existing models such as Voyage Code 3, Cohere Embed v4.0, and OpenAI’s large embedding model. Its enhanced performance offers developers unmatched precision and speed for code retrieval tasks.
![Codestral Embed]()
Customizable Dimensions and Precision for Efficiency
Offering embeddings in various dimensions and precisions, Codestral Embed allows users to balance retrieval quality against storage costs. For example, a 256-dimensional embedding with int8 precision still outperforms competitor models. Dimensions are ranked by relevance, enabling users to select the top n dimensions to optimize both quality and cost.
Benchmark Results Highlight Practical Excellence
![Results Highlight]()
Codestral Embed delivers outstanding results across multiple benchmark categories.
- SWE-Bench Lite: Evaluates real-world GitHub issues and fixes, key for retrieval-augmented generation in coding assistants.
- Text2Code (GitHub): Measures performance on context retrieval for code completion and editing.
Across these categories, Codestral Embed consistently outperforms competitors, proving its effectiveness for AI-driven code assistance.
Versatile Use Cases Driving Developer Productivity
Codestral Embed supports a wide range of practical applications within software development workflows.
- Retrieval-Augmented Generation: Enables fast context retrieval for code completion, editing, and explanation, ideal for AI copilots and coding agents.
- Semantic Code Search: Provides accurate searches for relevant code snippets using natural language or code queries.
- Similarity Search & Duplicate Detection: Identifies near-duplicate or functionally similar code, aiding code reuse and licensing compliance.
- Semantic Clustering & Code Analytics: Facilitates unsupervised code grouping to analyze repositories, identify architecture patterns, and automate documentation.
Availability and Pricing Details
Codestral Embed is available via Codestral’s API under the name codestral-embed-2505 at $0.15 per million tokens. The batch API is offered at a 50% discount. For enterprises seeking on-premise deployment, Codestral’s applied AI team is available for consultation.
Developers can access detailed documentation and usage examples in Codestral’s cookbook to get started efficiently.
Recommended Chunking for Optimal Retrieval
For best results in retrieval applications, Codestral recommends chunking datasets into 3000-character segments with 1000-character overlaps. Using the full 8192-token context is possible but may reduce retrieval performance. Further chunking guidance is available in the cookbook.
Detailed Benchmark Overview
| Benchmark |
Description |
Category |
| SWE-Bench Lite |
Retrieve files needed to fix GitHub issues; critical for retrieval-augmented generation (RAG). |
swebench_lite |
| CodeSearchNet Code->Code |
Retrieve code snippets from GitHub projects that appear in the same context. |
code2code |
| CodeSearchNet Doc2Code |
Retrieve code from docstrings in GitHub repositories. |
text2code (github) |
| CommitPack |
Retrieve modified files based on commit messages. |
text2code (github) |
| Spider, WikiSQL, Synthetic Text2SQL |
Retrieve SQL queries from natural language inputs. |
text2sql |
| DM Code Contests, APPS, CodeChef, MBPP+ |
Match programming problems to solutions across algorithms and data science categories. |
text2code (algorithms/data science) |
Conclusion
Codestral Embed delivers a groundbreaking combination of precision, flexibility, and efficiency for code retrieval and semantic understanding, tailored for large-scale codebases and modern development workflows. It marks a significant advancement in AI-powered software engineering tools.