Getting started with Mistral

What Mistral is

Mistral AI builds language models and publishes many of them as open weights, which means you can download the model itself. There are two main products:

• Vibe is the assistant and coding agent. It runs on the web, on mobile, in your terminal and in your editor. You may still see the older name Le Chat, for example in the App Store.
• Studio is the developer console and API, where you create keys and try models in the Playground.

Choose a route

Prices change, so check Mistral’s pricing page before you commit.

Vibe on the web

Go to chat.mistral.ai, create an account and start typing. The Free plan includes limited messages and web searches. There is nothing to install.

Vibe in the terminal

Vibe’s command-line tool is a coding agent that works in your project folder. Commit your work first, so that anything it changes is easy to undo.

curl -LsSf https://mistral.ai/vibe/install.sh | bash

uv tool install mistral-vibe

Then start it from your project folder:

vibe

The first launch runs a setup wizard. Choose Launch browser to sign in, or Use an API key. Run vibe --setup to go through it again. On Windows, install uv first, then use the uv tool install line. The install docs have the details.

Your first API call

1. Create an account at console.mistral.ai. Free mode is on by default and needs no credit card, though usage and rate limits apply.
2. Open Studio, then API Keys, and click Create new key.
3. Name the key, set an expiration date and copy it straight away. The full key appears only once.
4. Put the key in an environment variable, not in your code.

export MISTRAL_API_KEY="your_api_key_here"

$env:MISTRAL_API_KEY="your_api_key_here"

Then send a request:

curl https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-small-latest",
    "messages": [{"role": "user", "content": "Hello, Mistral!"}]
  }'

For Python, run pip install mistralai, save this as quickstart.py and run python quickstart.py:

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": "user", "content": "What is Mistral AI?"}
    ],
)

print(response.choices[0].message.content)

A 402 error means the account has no payment method. You can add one in billing settings.

Which local model fits your computer

Start with Ministral 3. Mistral designed this family for running on your own hardware, and it accepts text and images. All three sizes are Apache 2.0 licensed.

Download sizes come from Ollama’s library, which lists a maximum context of 256K for Ministral 3 and 384K for Devstral Small 2. The memory column is our own rough starting point, not a vendor figure. The model has to fit in memory alongside everything else you have open, so go one size down if it struggles. Mistral’s one stated figure is that Ministral 3 8B fits in 12 GB of graphics memory at FP8, and in less when quantized further.

A GPU helps a lot, but it is not required. Apple Silicon Macs share memory between CPU and GPU, so count your total. On a PC with an Nvidia card, count the card’s memory for full speed. A model that spills over into system memory still runs, only slower.

Install and run with Ollama

Ollama downloads and runs models for you from one command. Install it with the line for your system, or use the installer from ollama.com/download if you would rather not pipe a script into your shell.

curl -fsSL https://ollama.com/install.sh | sh

irm https://ollama.com/install.ps1 | iex

Then open a new terminal and start the model. The first run downloads it, and after that you can chat with no internet connection.

ollama run ministral-3:8b

Type a message to chat, and /bye to leave. These commands help you keep track of what is on your machine:

In ollama ps, the Processor column shows 100% CPU when the model is running only in system memory. If you have a GPU and expected it to be used, check your driver: Ollama needs Nvidia driver 550 or newer. See Ollama’s hardware support page for other cards.

/set parameter num_ctx 8192

Prefer an app? Use LM Studio

LM Studio does the same job with a graphical interface and no terminal.

1. Install it and open the Discover tab.
2. Search for ministral-3, then download the 3B, 8B or 14B version.
3. Open the Chat tab, choose the model in the model loader and start typing.

LM Studio recommends at least 16 GB of RAM. On Windows it also recommends 4 GB of dedicated graphics memory and needs a CPU with AVX2. On a Mac it needs Apple Silicon and macOS 14 or newer, and on an 8 GB Mac you should stick to smaller models. Check the full requirements before you download anything large.

Call the local model from your code

While Ollama is running, it serves an OpenAI-compatible API on your own machine. No key is needed, and this request never leaves your computer:

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ministral-3:8b",
    "messages": [{"role": "user", "content": "Hello from my own computer"}]
  }'

Tools that let you set an OpenAI base URL can use http://localhost:11434/v1/. Remember that the 4,096-token default applies here as well, so see Ollama’s FAQ for how to change it through the API.

What local will not do

• Mistral’s biggest models will not fit. Large 4 has about 1 trillion parameters and Small 4 has 119 billion. Despite its name, Small 4 is a server model, and Mistral points to serving software such as vLLM for those.
• Cloud tags are not local. Ollama lists mistral-large-4 as a cloud model. It runs on someone else’s servers, so your prompts leave your machine.
• Smaller models are less capable. Expect a local Ministral to fall short of the top hosted models on hard coding and reasoning work.
• Licences differ between models. Ministral 3’s model card allows commercial use under Apache 2.0, but other models have their own terms. Read the licence on the model’s card before using one in a business.

What to read next

See what Mistral has released lately on Latest updates, and compare it with other vendors on the models page. If you are still choosing a coding tool, read Getting started with Claude Code or Getting started with Cursor.