Overview
Codestral was Mistral AI's first dedicated code model. With 22B parameters and a 32K context window, it hit a sweet spot between quality and hardware requirements, and it was quickly integrated into editor plugins such as Continue and Tabnine.
Strengths
- Excellent fill-in-the-middle performance for inline completion
- 32K context, long enough for repository-level completion
- Fits on a single 24 GB GPU with 4-bit quantization
Run it locally
ollama run codestralLicense note
The open weights are under the Mistral AI Non-Production License. Check the license before using it in a product.
Good fit for
Personal and research use, and evaluating code completion quality before committing to a commercial license.
Frequently asked questions
Can I use Codestral in a commercial product?
Not with the open weights. The Mistral AI Non-Production License allows research and testing only. Commercial use requires a separate license or Mistral's API.
Is Codestral good for autocomplete?
Yes. It was designed with fill-in-the-middle in mind and is a popular choice for IDE autocomplete.