Mistral free AI: Vibe (formerly Le Chat) or local ?

#Mistral AI for free: the short answer

Mistral AI is a French company with two lines of business. It operates a consumer-facing online assistant on its servers, with a free plan and subscriptions. It also publishes the weights of some of its models on Hugging Face, under licenses that allow anyone to download and run them at home. Both paths are free, but not in the same way.

Free online?

Yes: the Mistral assistant offers a subscription-free plan, available after creating an account. It is subject to usage limits set by the publisher.

Unlimited?

No. The limits depend on the function used and the plan you subscribed to. They are described in the official help and change over time, so we do not reproduce any of them here.

Free to run locally?

Yes, with no ceiling: several Mistral models are released under the Apache 2.0 license and can be run with Ollama. The real cost is your hardware and electricity.

The same model on both sides?

Not necessarily. The online service runs on servers sized for large models. On a typical PC, you run models from the family's 7 to 24 billion parameter range.

Which one should you choose?

The online assistant for low-stakes questions and ready-made features. Go local whenever the text must not leave your machine or you want uncapped usage.

The rest covers each point in this order. If local execution is the only topic you care about, go straight to the section on the Mistral models to run locally.

#Le Chat, Vibe, Mistral Vibe: three names to untangle

The terminology changed in 2026, and most online tutorials have not caught up. According to Mistral’s official help article titled “Le Chat is now Vibe,” the consumer assistant has been called Vibe since May 28, 2026. We recorded this information from Mistral’s help center on September 29, 2026; the article’s URL appears at the end of the guide. The service changed its name, not its nature: it is still Mistral’s conversational assistant.

Le Chat

The former name of Mistral's online assistant. You'll still see it in articles, videos, and comparisons from before summer 2026.

Vibe (the assistant)

The current name of the same assistant, according to the official help. This is the one meant when discussing Mistral's free offering for the general public.

Mistral Vibe (the CLI)

A different tool: a code agent used in a terminal, introduced by Mistral in its announcement “Mistral Vibe 2.0”. It is aimed at developers and is not the subject of this guide.

L'API

Programmatic access to models for developers integrating Mistral into their own software. It has its own pricing and limits, separate from those of the assistant.

Open models

The weight files published by Mistral (Mistral Small, Mistral Nemo, Magistral, Devstral...). No Mistral account is required to run them.

#What Mistral AI’s free online offering includes

The free offering provides access to the assistant from a browser or the provider's mobile app after you create an account. You ask your questions in French, and the model answers in French: Mistral trains its models on many languages, and there is no particular “French” version to look for. Beyond conversation, the assistant offers additional features (web search, file reading, image generation, among others), whose availability and limits depend on the offering.

1. 01
   Start from the official website

   Type mistral.ai yourself and follow the link to the assistant. This helps you avoid third-party websites and apps that reuse the Mistral name while adding their own data collection, and sometimes their own subscription.
2. 02
   Create an account

   Sign-up requires an email address or a third-party account. The free plan does not require a subscription. Nothing requires you to use your work address.
3. 03
   Open the settings before your first real use

   Find your account's privacy page: that's where, depending on your plan, you determine whether your conversations can be used to improve the models. Make this choice before pasting in your first document, not afterward.
4. 04
   Write in French and test on your tasks

   Pick three real-world tasks without sensitive data (an email to rewrite, a text to summarize, and a question from your field). Ten minutes is enough to find out whether the free offering meets your needs.

https://mistral.ai

#Free-tier limitations

Free doesn't mean unlimited. Like all online assistants, Mistral's assistant caps usage on free accounts and offers subscriptions to go beyond that. What matters is the nature of these limits more than their current value.

Limits by function

Simple chat, web search, file analysis, and image generation do not cost the provider the same and are not capped in the same way. Reaching a limit on one feature does not necessarily block the others.

Rolling caps

As a general rule, a reached limit is lifted after a period set by the publisher. You often discover it only when you reach it because no counter is displayed in advance.

Served model

The provider decides which model responds and can change it. With a free plan, you control neither the version nor how long it remains available.

No service guarantee

A free offering comes with no availability commitment. If a job depends on it, this is not the right support.

https://help.mistral.ai/

#And the Mistral API, and Mistral OCR?

These are developer products, separate from the assistant. The API is controlled with a key and measured in tokens consumed; its trial terms and pricing are published in the Mistral documentation and cannot be inferred from the assistant's free plan. Mistral OCR, the document text-extraction service, is part of this API; to our knowledge, it is not a model offered for free download. If you need to read PDFs or invoices on your machine, local tools are available; they are covered in other guides on the site.

#What happens to your text in the online assistant

Mistral AI is a French company subject to the GDPR. That is a real advantage over a service based outside the European Union: you exercise your rights of access, rectification, and deletion with a provider directly subject to European law. That does not change the basic fact. Everything you write in an online assistant leaves your device, is processed on the provider's servers, and is stored there according to its rules.

Model training

Depending on the plan, conversations may be used to improve the models. The corresponding setting is in your account’s privacy settings: check what it says for your plan rather than assuming.

Retention

Deleting a conversation removes it from your history. Server-side retention periods are those specified in Mistral's privacy policy, whose revision date appears at the top of the document.

Uploaded files

A document sent for analysis is transmitted to the service in the same way as a message. A PDF containing names, addresses, or amounts remains personal data once uploaded.

Professional use

Offers for teams and businesses come with contractual commitments that differ from those of a free account. For customer or employee data, this is a decision to validate internally, not an individual choice.

“French AI” therefore does not mean “nothing leaves my environment.” The only configuration in which that statement is true is one where the model runs on your own machine. Mistral makes precisely that possible, which is rare among consumer-assistant vendors.

#Online or local: how to decide

The choice depends on three criteria: what the text contains, how regularly you use it, and the hardware you have. Neither approach is inherently better.

One-off questions, general knowledge, risk-free rewording

The free online assistant. Nothing to install, plus ready-to-use features such as web search.

Intensive daily use

You will eventually hit the limits of the free tier. Two options: a subscription, or a local model whose only limit is your hardware.

Personal or confidential documents

Local. Contracts, medical records, accounting, customer data: a model running on your machine sends nothing.

Workstation without a graphics card, older computer

The online assistant, or a small local model if you accept slow responses. A 7-billion-parameter model remains usable on the CPU alone, but nothing more without patience.

Need for stable behavior over time

Local. A downloaded weights file does not change until you replace it, whereas an online service evolves according to the schedule chosen by its publisher.

No reliable connection

Local. Once the model is downloaded, it runs offline.

In practice, the two complement each other: the online assistant for anything that could be read by anyone, and a local model for the rest. Starting with the free online option also lets you verify, without installing anything, that you like the response style of the Mistral models.

#The Mistral models to run at home

Mistral publishes its open models on its Hugging Face page, and the Ollama library includes most of them in ready-to-use form. Once the file is downloaded, the model runs on your graphics card or processor: no account, no cap, no data transmission. Here are the family’s models that fit on a personal machine, along with the VRAM required for Q4_K_M quantization.

Mistral 7B Instruct (mistral)

7 billion parameters, Apache 2.0 license, about 5 GB of VRAM. The landmark model from 2023: surpassed in quality, but it starts on almost any machine.

Mistral Nemo 12B (mistral-nemo)

12 billion parameters, Apache 2.0, about 7 GB of VRAM. A good starting point on a 12 GB card such as a RTX 3060 or RTX 4070.

Mistral Small 3.2 24B (mistral-small3.2)

24 billion parameters, Apache 2.0, about 14 GB of VRAM. General-purpose model with vision. It requires a 16 GB card (RTX 4080) or 24 GB card (RTX 4090), or a Mac with 24 GB or more of unified memory.

Magistral Small 24B (Magistral)

24 billion parameters, Apache 2.0, approximately 14 GB of VRAM. The reasoning variant, for multi-step problems. It has its own installation guide on the site.

Devstral Small 2 24B (devstral-small-2)

24 billion parameters, Apache 2.0, about 14 GB of VRAM. Specialized in code and development agents.

These memory values are the site's catalog sizing benchmarks for Q4_K_M quantization. They are not speed measurements: throughput depends on your hardware and context length, and we do not provide it here. On a 12 GB card, a 24-billion-parameter model spills into system RAM and slows down significantly; it is better to stick with Mistral Nemo.

Mistral also publishes the weights of much larger models, such as Mistral Small 4 (119 billion parameters despite its name), Mistral Medium 3.5, or Mistral Large 3. They belong on a workstation or server, with several dozen to several hundred gigabytes of memory, and fall outside the scope of a first free trial.

1. 01
   Install Ollama

   Ollama is the program that downloads and runs models. It is available for Windows, macOS, and Linux and can be downloaded from ollama.com. Once started, it listens on http://localhost:11434.
2. 02
   Choose the model based on your graphics memory

   Up to 8 GB of VRAM: Mistral 7B. With 12 GB: Mistral Nemo. From 16 GB: Mistral Small 3.2. When in doubt, choose the smaller size: a model that fits entirely in VRAM is much more pleasant to use than a larger model that spills over.
3. 03
   Run the model

   A single command downloads the model on first launch and then opens the conversation in the terminal. Write in French, as you would with the online assistant.
4. 04
   Add an interface

   To recover a chat window with history, Open WebUI or LM Studio connect to the local model. Everything stays on your machine.

# Carte de 12 Go : Mistral Nemo 12B
ollama run mistral-nemo

# Carte de 16 Go et plus : Mistral Small 3.2 24B
ollama run mistral-small3.2

# Machine modeste : Mistral 7B
ollama run mistral

To verify that the response really comes from your machine, query the Ollama service directly. The localhost address never leaves your computer, and the command works with the network disconnected once the model has been downloaded.

curl http://localhost:11434/api/generate -d '{
  "model": "mistral-nemo",
  "prompt": "Réponds en français : explique en trois phrases la licence Apache 2.0.",
  "stream": false
}'

#Pitfalls and misconceptions

“Vibe is the tool for coding”

The name refers to two things. The consumer assistant has been called Vibe since May 2026; the command-line coding agent is called Mistral Vibe. Check whether the article you’re reading is discussing a terminal or a chat window.

“Locally, I have exactly the same assistant as online”

No. You get a model, not a service: no built-in web search, no image generation, and no synchronization between devices unless you add them yourself. And the model that fits on a PC is smaller than those a provider can serve from its data centers.

“Small means small”

With Mistral, the name refers to a lineup, not a size. Mistral Small 3.2 has 24 billion parameters, while Mistral Small 4 has 119. Go by the parameter count and file size, never the name.

“French AI, so my data is safe”

The legal framework is European, which matters. But text sent to an online service is still text sent to an online service. Strict confidentiality is achieved locally, regardless of the publisher's country.

“The quota I read about on a blog is reliable”

It may have been on the day it was published. Only Mistral's help center describes the limits currently in effect.

“Free locally, so there’s no cost”

The software and weights are free. The hardware is not: a 12 GB graphics card is the comfortable entry point, and electricity adds to the cost for sustained use.

#Official sources to consult

The information in this guide about the online assistant refers to Mistral’s pages, reviewed on September 29, 2026. They may change without notice: if they differ from this guide, the official page takes precedence.

https://help.mistral.ai/en/articles/682992-le-chat-is-now-vibe

https://docs.mistral.ai/getting-started/models/

https://mistral.ai/news/mistral-vibe-2-0/

https://huggingface.co/mistralai

For the local component, each model page on Hugging Face lists its license, size, and publication date. This is the reference to consult before any professional use.

#Go further

This guide helps you choose between the online assistant and a local setup. To put it into practice, these site guides take over:

Install Magistral

The step-by-step installation of Mistral’s reasoning model with Ollama, including configuration. https://quelllm.fr/guide/mistral-magistral-installation

Free local AI

The free tools and models to install based on your machine's memory, across all families. https://quelllm.fr/guide/ia-locale-gratuite

Best local LLM for coding

Devstral vs. Qwen3-Coder: which coding model to choose for your graphics card. https://quelllm.fr/guide/meilleur-llm-local-pour-coder-2026-devstral-qwen3-coder

Local AI in business and GDPR

The framework to establish before deploying an open model such as Mistral Small for a team. https://quelllm.fr/guide/ia-locale-entreprise-rgpd