Overview

Mistral AI is a French AI company whose API serves chat, reasoning, code (Codestral), embedding, OCR and speech models. The chat endpoint (/v1/chat/completions) follows OpenAI conventions. Requests go to api.mistral.ai by default; EU or US processing is available through paid regional endpoints (see Regional inference). The official SDKs are mistralai (2.x and 3.x; 3.3.0 is current) for Python and @mistralai/mistralai for TypeScript.

Setup

bash

# Python 3.10+
pip install mistralai

# TypeScript/Node
npm install @mistralai/mistralai

bash

export MISTRAL_API_KEY=...

Available Models

Model (alias)	Context	Best for
`mistral-large-latest`	256k	Large 3, flagship general model (Large 4 is `mistral-large-4`, 1M context)
`mistral-medium-latest`	256k	Medium 3.5, multimodal, agentic and coding work
`mistral-small-latest`	256k	Small 4, cheapest everyday chat, reasoning and code
`codestral-latest`	128k	Code generation and fill-in-the-middle
`mistral-embed`	8k	Text embeddings (1024 dimensions)

Aliases move to newer versions; pin a dated name (mistral-small-2603, codestral-2508) in production. open-mistral-nemo was deprecated in May 2026 with retirement scheduled for 2026-07-31 (docs.mistral.ai/models still lists it as Deprecated). The full list with status and retirement dates is at docs.mistral.ai/models.

Instructions

Basic Chat Completion (Python)

In mistralai 2.x and later the import is from mistralai.client import Mistral; from mistralai import Mistral raises ImportError. On the old 1.x line pin pip install "mistralai<2".

python

import os
from mistralai.client import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

response = client.chat.complete(
    model="mistral-large-latest",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain the difference between async and sync programming."},
    ],
)

print(response.choices[0].message.content)
print(f"Prompt tokens: {response.usage.prompt_tokens}")
print(f"Completion tokens: {response.usage.completion_tokens}")

TypeScript/Node.js

typescript

import { Mistral } from "@mistralai/mistralai";   // named export; there is no default export

const client = new Mistral({ apiKey: process.env.MISTRAL_API_KEY });

const response = await client.chat.complete({
  model: "mistral-large-latest",
  messages: [{ role: "user", content: "Hello from TypeScript!" }],
});

console.log(response.choices[0].message.content);

Streaming

python

from mistralai.client import Mistral

client = Mistral()

stream = client.chat.stream(
    model="mistral-small-latest",
    messages=[{"role": "user", "content": "Write a haiku about programming."}],
)

for event in stream:
    chunk = event.data.choices[0].delta.content
    if chunk:
        print(chunk, end="", flush=True)
print()

Function Calling

python

import json
from mistralai.client import Mistral

client = Mistral()

tools = [
    {
        "type": "function",
        "function": {
            "name": "search_products",
            "description": "Search for products in a catalog",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string"},
                    "max_price": {"type": "number"},
                    "category": {"type": "string"},
                },
                "required": ["query"],
            },
        },
    }
]

messages = [{"role": "user", "content": "Find laptops under 1000 dollars"}]

response = client.chat.complete(
    model="mistral-large-latest",
    messages=messages,
    tools=tools,
    tool_choice="auto",   # default; "any" forces a call, "none" forbids one
)

if response.choices[0].finish_reason == "tool_calls":
    tool_call = response.choices[0].message.tool_calls[0]
    args = json.loads(tool_call.function.arguments)
    print(f"Function: {tool_call.function.name}, Args: {args}")

    # Add tool result and continue
    messages.append(response.choices[0].message)   # keep the assistant message with its tool_calls
    messages.append({
        "role": "tool",
        "tool_call_id": tool_call.id,
        "name": tool_call.function.name,
        "content": json.dumps([{"name": "ThinkPad X1", "price": 899}]),
    })

    final = client.chat.complete(model="mistral-large-latest", messages=messages)
    print(final.choices[0].message.content)

JSON Mode

python

from mistralai.client import Mistral
import json

client = Mistral()

response = client.chat.complete(
    model="mistral-small-latest",
    messages=[
        {
            "role": "user",
            "content": "Return a JSON object with fields: title, author, year for the book '1984'",
        }
    ],
    response_format={"type": "json_object"},
)

data = json.loads(response.choices[0].message.content)
print(data)  # {"title": "1984", "author": "George Orwell", "year": 1949}

JSON mode only guarantees valid JSON, not your field names. For a guaranteed schema use structured outputs: define a Pydantic model and call client.chat.parse(model=..., messages=..., response_format=Book); the parsed object is on response.choices[0].message.parsed.

Text Embeddings

python

from mistralai.client import Mistral

client = Mistral()

response = client.embeddings.create(
    model="mistral-embed",
    inputs=["Machine learning is transforming industries.", "AI is the future of technology."],
)

embeddings = [item.embedding for item in response.data]
print(f"Embedding dimension: {len(embeddings[0])}")  # 1024

Codestral for Code Completion

python

from mistralai.client import Mistral

client = Mistral()

# Fill-in-the-middle (FIM) — Codestral's signature feature
response = client.fim.complete(
    model="codestral-latest",
    prompt="def fibonacci(n):\n    if n <= 1:\n        return n\n    ",
    suffix="\n\nresult = fibonacci(10)\nprint(result)",
)

print(response.choices[0].message.content)
# Returns the middle code that connects prompt to suffix

Examples

Example 1: Summarise support tickets in French with EU processing

Request: "Summarise these three support tickets in French and keep inference in the EU."

python

import os
from mistralai.client import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"], server="eu")
tickets = "1) Facture en double\n2) Mot de passe oublie\n3) Livraison en retard"
r = client.chat.complete(
    model="mistral-small-latest",
    messages=[{"role": "user", "content": f"Resume ces tickets en une phrase chacun:\n{tickets}"}],
)
print(r.choices[0].message.content)
print(r.usage.total_tokens)

Result: three one-line French summaries from api.eu.mistral.ai, and a token count to track cost at the 1.1x regional rate.

Example 2: Complete a function with Codestral

Request: "Fill in the body of this function between the signature and the test call."

python

r = client.fim.complete(
    model="codestral-latest",
    prompt="def slugify(title: str) -> str:\n    ",
    suffix="\n\nprint(slugify('Hello, World!'))",
)
print(r.choices[0].message.content)

Result: only the missing middle (for example return re.sub(r"[^a-z0-9]+", "-", title.lower()).strip("-")), ready to splice between prompt and suffix. Add import re yourself if the model does not.

Regional inference and GDPR

The default endpoint api.mistral.ai makes no commitment about where requests are processed. For EU processing use the regional endpoint, billed at 1.1x list price:

python

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"], server="eu")   # api.eu.mistral.ai; TypeScript: { server: "eu" }

Regional endpoints serve only models hosted in that region (check with client.models.list()), support function calling as the only tool, and do not offer Agents, Batch or the Files API. Account data, API keys and billing metadata are not regional. Zero data retention and data-processing terms are arranged with Mistral for your organization; read the DPA rather than assuming a default.

Guidelines

Use mistral-large-latest or mistral-medium-latest for complex tasks, mistral-small-latest for cost savings.
Codestral is specialised for code and for fill-in-the-middle (client.fim.complete).
The mistral-embed model produces 1024-dimensional vectors.
Mistral models have strong multilingual performance, especially in French, Spanish, Italian, German, and Portuguese.
tool_choice defaults to "auto"; always send the tool result back with tool_call_id.
In JSON mode also ask for JSON in the prompt and name the fields you expect.
Read the API key from MISTRAL_API_KEY; never hard-code it.
Aliases such as *-latest change under you; pin dated model names for reproducible output.

Mistral AI API

Validation

Getting Started

Example Prompts

Documentation

Overview

Setup

Available Models

Instructions

Basic Chat Completion (Python)

TypeScript/Node.js

Streaming

Function Calling

JSON Mode

Text Embeddings

Codestral for Code Completion

Examples

Example 1: Summarise support tickets in French with EU processing

Example 2: Complete a function with Codestral

Regional inference and GDPR

Guidelines

Information

Use Cases

Related Skills