← Labs & models

Lab

Mistral AI.

Paris, FranceFounded 2023Models: Some online, some downloadable

In plain English

Mistral is a French AI company and one of Europe’s best-known AI labs. Its stated mission is to make advanced AI accessible to everyone. It presents itself as a European alternative to US and Chinese providers, with an emphasis on giving organisations more control over where and how they use AI.

Vibe, formerly called Le Chat, is Mistral’s AI assistant for chat, work, and coding. It is the product you use, in much the same way as ChatGPT is an OpenAI product. Behind it are Mistral’s models. Some are open weight, which means you can download and customise them, subject to their licence terms.

Mistral also puts emphasis on multilingual AI. It says Mistral Large 4 was trained on data covering more than 160 languages, including every official EU language.

Mistral also offers OCR, a document-reading service that extracts and structures information from files.

Models to know

Mistral Large 4 for demanding work, including coding and tasks involving different types of information. Mistral Small is a smaller open-weight model that can be downloaded and customised.

Who's in charge

Arthur Mensch, co-founder and CEO.

Things to be aware of

Mistral says data on its services is hosted in the EU by default, but some features can temporarily transfer data outside the EU. Check the feature, settings, and contract rather than assuming everything stays in Europe.

Vibe’s data-training settings also depend on the plan. Individual users are not opted out by default, while Enterprise customers are. Check the settings before your team uploads company information.

If you use a Mistral model through another company’s cloud, or download and run it yourself, where the data goes depends on that setup. Choosing a European lab does not automatically make your use GDPR-compliant or remove any obligations under the EU AI Act.

Read more

What happened to Le Chat? · Mistral data location · Mistral data-training settings

3 model cards

Mistral Large 4.

Power

££

API cost

In plain English

Mistral’s most powerful model. You can use it online or download the weights to run it yourself.

Good for

Coding, working with documents and images, multilingual tasks, and organisations that want more control over where their AI runs.

Things to be aware of

Downloading the model gives you more control, but you also take on the infrastructure, security, support, and running costs. Check where it runs and what your contract says before claiming your data stays in Europe.

The licence allows many commercial uses, but businesses offering it commercially as a model service need a separate licence from Mistral.

API price
$0.70 in / $2.10 out
per million tokens
Context window
1.0M tokens

Mistral Small 4.

Power

££

API cost

In plain English

A smaller Mistral model for everyday work. It can handle text and images, and you can download and adapt it yourself.

Good for

Chat assistants, document understanding, coding, and repeated tasks where cost and speed matter.

Things to be aware of

“Small” is relative. Running it yourself still needs substantial computing infrastructure. Test it on the work you need done before deciding that the lower price is a saving: checking and redoing weak results takes time too.

API price
$0.15 in / $0.60 out
per million tokens
Context window
256.0K tokens

Voxtral.

Specialist

Power

Varies by model

API cost

Model family

In plain English

Mistral’s family of speech models. Some turn speech into text. Another turns text into speech.

Good for

Transcribing meetings and calls, adding voice to an application, and building voice assistants.

Things to be aware of

Transcripts can get words or speakers wrong, so check important details. Voice generation also raises questions about consent and impersonation. Be clear about whose voice can be used and how.

What Mistral says

“Speech to text turns recorded files and live audio into text”
Mistral’s speech-to-text documentation