Codestral is Mistral AI’s dedicated coding model — a 22-billion-parameter model trained from the ground up on 80+ programming languages. Unlike general-purpose models that happen to code, Codestral is purpose-built for code generation, completion, and fill-in-the-middle (FIM) tasks.

It’s the best autocomplete model available and one of the most efficient coding models you can run locally.

Why Codestral matters

Most AI coding happens through two workflows:

1. Chat/agent — describe what you want, AI writes it (Claude Code, Aider)
2. Autocomplete — AI predicts what you’ll type next (Copilot, Cursor tab)

Codestral dominates the second category. Its FIM (Fill-in-the-Middle) capability means it understands code before AND after your cursor, producing completions that fit naturally into existing code.

Specs

Spec	Codestral 25.01
Parameters	22B
Architecture	Dense transformer
Context window	256K tokens
Languages	80+
FIM support	✅ Native
HumanEval	86.6%
License	Mistral Non-Production License
Quantized size	~12GB (Q4)

The 256K context window is massive for a 22B model — it can see your entire codebase in a single pass.

Benchmarks

Benchmark	Codestral 25.01	DeepSeek Coder 33B	CodeLlama 70B
HumanEval	86.6%	79.3%	67.8%
MBPP	78.2%	73.1%	62.0%
RepoBench (long-range)	Best	Good	Limited

Codestral outperforms models 3x its size on code generation benchmarks. The RepoBench score is particularly impressive — it handles long-range code completion across entire repositories better than any model in its class.

How to use Codestral

Via Mistral API

from mistralai import Mistral

client = Mistral(api_key="your-mistral-key")

# Chat completion
response = client.chat.complete(
    model="codestral-latest",
    messages=[{"role": "user", "content": "Write a binary search in Rust"}]
)

# Fill-in-the-middle
response = client.fim.complete(
    model="codestral-latest",
    prompt="def fibonacci(n):\n    ",
    suffix="\n    return result"
)

Via OpenRouter

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="your-openrouter-key"
)

response = client.chat.completions.create(
    model="mistralai/codestral-latest",
    messages=[{"role": "user", "content": "Optimize this SQL query"}]
)

See our OpenRouter guide for setup.

Run locally with Ollama

Codestral at 22B fits comfortably on consumer hardware:

ollama pull codestral:22b
ollama run codestral:22b

Hardware	Speed	Usable?
RTX 4090 (24GB)	~40 tok/s	✅ Excellent
RTX 4070 (12GB)	~20 tok/s (Q4)	✅ Good
Mac M4 32GB	~25 tok/s	✅ Good
Mac M4 16GB	~15 tok/s (Q4)	⚠️ Tight

See our Ollama guide for detailed setup.

In VS Code with Continue.dev

Codestral is one of the best models for Continue.dev autocomplete:

{
  "tabAutocompleteModel": {
    "provider": "ollama",
    "model": "codestral:22b",
    "title": "Codestral"
  }
}

Codestral vs other coding models

	Codestral 22B	Gemma 4 27B	Qwen 2.5 Coder 32B	DeepSeek Coder 33B
Best at	Autocomplete/FIM	General + coding	Coding breadth	Reasoning + coding
FIM	✅ Native	❌	✅	✅
Context	256K	128K	128K	128K
Local VRAM	12GB (Q4)	16GB (Q4)	18GB (Q4)	18GB (Q4)
License	Non-production	Gemma	Apache 2.0	MIT

Pick Codestral for: IDE autocomplete, FIM tasks, fast local completions. Pick Gemma 4 for: General-purpose coding + other tasks. Pick Qwen Coder for: Broadest language support, commercial use.

Licensing caveat

Codestral uses Mistral’s Non-Production License (MNPL), which means:

• ✅ Free for research and personal use
• ✅ Free for development and testing
• ❌ Cannot be used in production commercial applications
• ❌ Cannot be redistributed

For commercial production use, you need Mistral’s commercial license or should use Qwen 2.5 Coder (Apache 2.0) or DeepSeek Coder (MIT) instead.

Bottom line

Codestral is the best autocomplete model available. At 22B parameters, it runs on consumer hardware while outperforming models 3x its size. The 256K context window means it understands your entire project. If you’re setting up a local AI coding environment, Codestral should be your autocomplete model — paired with a larger model like Claude or GLM-5.1 for complex tasks.

FAQ

Is Codestral free?

Yes, for non-commercial use. Codestral is free for research, personal projects, and development/testing under Mistral’s Non-Production License. For production commercial use, you need a commercial license from Mistral. See our full Codestral overview for licensing details.

How does Codestral compare to Copilot?

Codestral is an open-weight model you can run locally or via API, while Copilot is a closed SaaS product from GitHub/Microsoft. Codestral gives you more control, no subscription cost, and works offline — but Copilot has tighter IDE integration out of the box. For a broader look, see our AI model comparison.

Can I run Codestral locally?

Yes. At 22B parameters (~12GB quantized to Q4), Codestral runs well on consumer GPUs like the RTX 4070/4090 or Apple Silicon Macs with 16GB+ RAM. Use Ollama or llama.cpp to run it. See our best AI models for coding locally guide for hardware recommendations.

What’s the difference between Codestral and Mistral Large?

Codestral (22B) is purpose-built for code — it has native FIM support, faster inference, and lower hardware requirements. Mistral Large is a general-purpose model that handles coding alongside other tasks but lacks FIM and is too large to run locally. Pick Codestral for autocomplete and code generation; pick Mistral Large for mixed workloads that include reasoning, writing, and code.

Related: What Is Codestral? · Codestral vs DeepSeek Coder · Best AI Models for Coding Locally · AI Model Comparison · Continue.dev Complete Guide