Curiosity · AI Model

Codestral

Codestral is Mistral AI's 2024–2025 code-specialized open-weights family — the latest Codestral 25.01 refresh delivers best-in-class fill-in-the-middle completion, 80+ language coverage, and competitive HumanEval scores. Weights are public under the Mistral Non-Production License (MNPL) with a commercial path via the API.

Model specs

Vendor
Mistral AI
Family
Codestral
Released
2025-01
Context window
256,000 tokens
Modalities
text, code
Input price
$0.3/M tok
Output price
$0.9/M tok
Pricing as of
2026-04-20

Strengths

  • Open weights released for research and non-production use
  • Industry-leading fill-in-the-middle for IDE completion
  • 256K context window — works at repository scale
  • Broad language coverage — 80+ programming languages

Limitations

  • Mistral Non-Production License — not a true open-source license for commercial self-host
  • Commercial deployment requires Mistral's paid API or a separate license
  • Behind DeepSeek Coder V2 on some benchmarks despite newer release
  • Smaller than Qwen 2.5 Coder 32B but only marginally faster per-token

Use cases

  • Self-hosted IDE extensions with low-latency completion
  • Repository-scale refactoring with 256K context
  • Multi-language code generation and review pipelines
  • Fine-tuning base for language-specific coding assistants

Benchmarks

BenchmarkScoreAs of
HumanEval≈86%2025-01
MBPP≈81%2025-01
RepoBench≈38%2025-01

Frequently asked questions

What license is Codestral released under?

The Mistral Non-Production License (MNPL). Weights are downloadable for research, personal, and evaluation use, but production commercial use requires Mistral's paid API or a separate commercial license.

How does Codestral compare to DeepSeek Coder V2?

DeepSeek Coder V2 often scores higher on HumanEval and MBPP and is MIT-licensed. Codestral has better fill-in-the-middle behavior and longer context in the 25.01 release — pick based on license and use-case fit.

Is Codestral good for IDE completion?

Yes — fill-in-the-middle training makes it one of the strongest open models for IDE-style mid-line completion. Pair with a language server for best results.

Sources

  1. Mistral AI — Codestral 25.01 — accessed 2026-04-20
  2. Hugging Face — mistralai/Codestral-22B-v0.1 — accessed 2026-04-20