Mistral Codestral — 22B Open-Weight Coding Model

Quick Use

1. Local: ollama pull codestral && ollama run codestral
2. API: get a key at console.mistral.ai, then export MISTRAL_API_KEY=...
3. Plug into Continue / Aider / LiteLLM with the snippets below

Intro

Codestral is Mistral AI's coding-specialized model — 22B parameters, 32K context, trained on 80+ languages from Python and JavaScript to Bash and Fortran. The open-weight Codestral 22B is free for research and non-commercial use; the latest Codestral models are available via Mistral's API and Le Chat for production. Best for: local code completion via Continue / Aider, latency-sensitive code generation. Works with: Ollama, vLLM, llama.cpp, Continue, Aider, LiteLLM. Setup time: 5 minutes (Ollama pull or API key).

Run Codestral locally with Ollama

# Pull
ollama pull codestral

# Chat
ollama run codestral

# Use as completion server (port 11434)
curl http://localhost:11434/api/generate -d '{
  "model": "codestral",
  "prompt": "Write a Rust function that returns the Nth Fibonacci number using memoization."
}'

Use Codestral with Continue

// config.json in your Continue extension
{
  "models": [
    {
      "title": "Codestral (local)",
      "provider": "ollama",
      "model": "codestral",
      "apiBase": "http://localhost:11434"
    }
  ],
  "tabAutocompleteModel": {
    "title": "Codestral Autocomplete",
    "provider": "ollama",
    "model": "codestral"
  }
}

Use Codestral with Aider

# Via Mistral API (production)
export MISTRAL_API_KEY=your-key
aider --model mistral/codestral-latest

# Or via local Ollama
aider --model ollama/codestral

Through LiteLLM

from litellm import completion

response = completion(
    model="codestral/codestral-latest",
    messages=[{"role": "user", "content": "Refactor this Python function to use list comprehensions"}],
    api_key=os.environ["CODESTRAL_API_KEY"],
)

Codestral vs general Mistral models

FAQ

Q: Is Codestral free? A: The Codestral 22B open-weights are free under the Mistral AI Non-Production License (research, non-commercial). For production / commercial use, call Codestral via Mistral's paid API on la-platforme.mistral.ai.

Q: How does Codestral compare to Claude / GPT for coding? A: On code-specific benchmarks (HumanEval, MBPP) Codestral is competitive with Claude 3.5 Sonnet. For broad reasoning + coding combined, Claude / GPT-4 still lead. Codestral's wins are local deployment and lower latency.

Q: Can I fine-tune Codestral? A: Yes — Mistral provides LoRA fine-tuning recipes for Codestral. The open-weight version can be fine-tuned with any standard tool (axolotl, unsloth, trl). Mistral's hosted fine-tuning API is also available.

Source & Thanks

Built by Mistral AI. Open-weights under Mistral AI Non-Production License.

Codestral on Hugging Face — ⭐ 1,500+