Mistral Codestral Review — 22B Code Model, Fill-in-the-Middle, 32K Context

Mistral Codestral (May 29, 2024) is a 22.2B parameter code-specialized model with native fill-in-the-middle training, 32K context, and 80+ language support. HumanEval 81.1%, MBPP 78.2%, RepoBench 34.0% — outperforms Code Llama 70B and DeepSeek Coder 33B on HumanEval at less than a third the parameters. Restricted by Mistral AI Non-Production License. Ollama: codestral. Rating: 4/5.