Codestral Embed model release

TL;DR

Mistral AI announced Codestral Embed, its first embedding model specialized for source code, and demonstrated that it surpasses leading competitors (Voyage Code 3, Cohere Embed v4.0, OpenAI’s large embedding model) on real‑world code retrieval benchmarks.

Model Overview

Codestral Embed is designed for high‑performance code retrieval and semantic understanding. The model can emit embeddings of various dimensions (e.g., 256) and numeric precisions (including int8). Mistral reports that even the smallest configuration (256‑dimensional int8) outperforms every competing model they evaluated. The embedding dimensions are ordered by relevance, allowing users to truncate to any integer n and obtain a smooth trade‑off between retrieval quality and storage cost.

Benchmark Performance

Codestral Embed was benchmarked across multiple code‑centric retrieval tasks. The results are summarized by category, with detailed benchmark tables provided by Mistral.

Category	Representative Benchmarks
Retrieval‑augmented generation	SWE‑Bench lite (real GitHub issues → relevant files)
Semantic code search	CodeSearchNet (code→code, doc2code)
Similarity & duplicate detection	CommitPack (commit message → modified files)
Text‑to‑SQL	Spider, WikiSQL, Synthetic Text2SQL
Algorithmic code matching	DM code contests, APPS, CodeChef, MBPP+
Data‑science code matching	DS 1000

Mistral states that the model achieves the highest average scores in each category and leads the macro‑average across all categories.

Primary Use Cases

1. Retrieval‑augmented generation (RAG)

Codestral Embed enables rapid context retrieval for code completion, editing, or explanation, making it ideal for AI‑powered copilots and coding‑agent frameworks.

2. Semantic code search

The embeddings support accurate search of code snippets from natural‑language or code queries, suitable for integration into developer tools, documentation systems, and copilots.

3. Similarity search & duplicate detection

By capturing functional similarity despite lexical variation, the model helps identify reusable code, prevent duplicate implementations, and enforce licensing policies.

4. Semantic clustering & code analytics

Unsupervised grouping of code by functionality or structure becomes feasible, aiding repository composition analysis, architecture discovery, and automated documentation.

Availability & Pricing

• API name: codestral-embed-2505
• Cost: $0.15 per million tokens (standard API) — 50 % discount on the batch API.
• On‑premises: Available via direct contact with Mistral’s applied AI team.

Documentation, quick‑start guides, and a Colab cookbook are provided:

• Docs: https://docs.mistral.ai/capabilities/embeddings/code_embeddings/
• Cookbook: https://colab.research.google.com/github/mistralai/cookbook/blob/main/mistral/embeddings/code_embedding.ipynb

Recommended Chunking Strategy

For retrieval scenarios, Mistral advises chunking source files into ~3000‑character segments with a 1000‑character overlap. While the model can process up to 8192 tokens, larger chunks degrade retrieval performance.

Benchmark Details

Mistral lists the specific benchmarks used to evaluate Codestral Embed:

Benchmark	Description	Category
SWE‑Bench lite	Retrieve files to fix real GitHub issues	swebench_lite
CodeSearchNet (code→code)	Retrieve code appearing in the same context	code2code
CodeSearchNet doc2code	Retrieve code from a docstring query	Text2code (github)
CommitPack	Retrieve modified files from a commit message	Text2code (github)
Spider	Retrieve SQL code from a natural‑language query	Text2SQL
WikiSQL	Retrieve SQL code from a natural‑language query	Text2SQL
Synthetic Text2SQL	Retrieve SQL code from a synthetic query	Text2SQL
DM code contests	Match problem descriptions to correct competition solutions	Text2Code (Algorithms)
APPS	Match problem descriptions to competition solutions	Text2Code (Algorithms)
CodeChef	Match problem descriptions to competition solutions	Text2Code (Algorithms)
MBPP+	Match algorithmic questions to Python solutions	Text2Code (Algorithms)
DS 1000	Match data‑science questions to implementations	Text2Code (Data Science)

These benchmarks cover a broad spectrum of code retrieval tasks, confirming the model’s versatility across software engineering, data‑science, and algorithmic domains.

---

All statements are directly derived from Mistral AI’s official announcement dated 2025‑05‑28.