Before you start
- To understand how the models work and their limits, see Generative AI explained simply
In plain words
Mistral AI is a French company, founded in Paris in 2023, that designs its own AI models. It stands out in two ways: a large share of its models are open (you can download them and run them yourself), and it offers a European alternative to the American giants. Its assistant, long called Le Chat, has been called Vibe since May 2026: a single tool for office work and for code.
Information checked on 20 October 2026. No prices are given here.
1. Mistral in brief
| Element | Detail |
|---|---|
| Headquarters | Paris |
| Assistant | Vibe (formerly Le Chat): web (chat.mistral.ai), iOS, Android, VS Code and JetBrains extension, command line |
| Developer platform | Mistral Studio (console.mistral.ai), formerly "La Plateforme" |
| Models | A complete range, many of them under the Apache 2.0 licence |
| Infrastructure | Its own data centres in France: training site at Bruyères-le-Châtel, new inference site at Les Ulis (south of Paris), expansion in Sweden |
| Customers | Individuals, businesses (with private deployments), the public sector |
2. Vibe, the assistant
On 28 May 2026, Le Chat became Vibe. Conversations, settings and subscriptions were kept. Vibe is organised around two modes.
Work mode
For multi-step office work:
| Feature | What it does |
|---|---|
| Plan before acting | For a complex task, Vibe proposes a plan and waits for your approval before starting; each step and each tool used can be viewed |
| Connectors | More than 100 connectors: Google Workspace, Outlook, SharePoint, Slack, GitHub, and custom connectors |
| Scheduled tasks | A request run once, or every day, week or month, with a notification at the end |
| Canvas | Writing and editing a document in a dedicated space |
| Data | Analysing a spreadsheet or a database, producing charts |
| Skills | Reusable instructions to automate repetitive work |
| Projects and library | Grouping conversations and files by subject |
| Memory | Vibe can remember elements from previous conversations (can be turned off) |
| Voice | Voice mode based on the Voxtral models |
| Images | Creating and editing images |
The old simple conversation mode is being phased out: Le Chat's history remains available in Work mode.
Code mode
For development, Vibe offers three entry points, which share the same sessions:
| Tool | Use |
|---|---|
| Vibe CLI | Coding agent in the terminal, on Mistral's flagship models optimised for code and agents |
| Vibe for IDEs | VS Code and JetBrains extension: completion, edits, conversation |
| Vibe on the web | chat.mistral.ai/code: remote coding sessions, in isolated sandboxes, from feature to pull request, several in parallel |
The /teleport command moves a session from the terminal to the cloud, or the other way round: you start work on your machine and let it carry on remotely.
The plans
| Plan | What it brings |
|---|---|
| Free | Vibe on web and mobile with limits (messages, web searches, coding sessions), basic images, more than 100 connectors, 5 scheduled tasks |
| Pro (and Education for students) | Higher limits, unlimited scheduled tasks, 15 GB library, spreadsheets, slides, Canvas, coding all day in the CLI, IDE and web, remote coding agents |
| Team | Higher limits, shared workspace, 30 GB per user, domain verification, data export, administration |
| Enterprise | Private deployments, custom models and agents, audit logs, SAML single sign-on, admin API |
3. The models
| Model | Date | Licence | Use |
|---|---|---|---|
| Mistral Large 4 | October 2026 (public preview) | Open weights announced after safety testing | Mistral's largest model (around 1 trillion parameters), multimodal, 1 million-token context |
| Mistral Medium 3.5 | April 2026 | Modified MIT | Versatile flagship model (128 billion parameters), optimised for code and agents |
| Mistral Small 4 | March 2026 | Apache 2.0 | 119 billion parameters, text and images, excellent capability-to-cost ratio |
| Mistral Large 3 | December 2025 | Apache 2.0 | Large open model (mixture-of-experts architecture) |
| Ministral 3 | December 2025 | Apache 2.0 | Small models (3 to 14 billion parameters) for workstations, phones, devices |
| Codestral | 2025 | Proprietary | Fast code completion in the editor |
| Voxtral (Realtime, Transcribe, TTS) | 2025-2026 | Depends on the model | Transcription, voice conversation, speech synthesis |
| Mistral OCR | 2025-2026 | Proprietary | Extracting text and structure from scanned documents |
Names and versions change often: the 2025 Devstral (coding agents) and Magistral (reasoning) models have now been retired, their capabilities taken over by Medium 3.5 and the recent models. The up-to-date list is in the model documentation. In the API, aliases such as mistral-medium-latest always point to the latest version.
Open doesn't mean the same everywhere
An open-weights model can be downloaded, run and modified. But the licence matters: Apache 2.0 allows any use, including commercial; a modified MIT licence adds conditions (often for very large companies); a non-commercial licence (CC BY-NC) forbids resale. Read the licence before any professional use.
4. Mistral Studio and the API
Mistral Studio (console.mistral.ai) is the developers' console: trying models, API keys, usage tracking, agents, fine-tuning and batch processing. A free "Experiment" tier lets you test.
The API follows the same format as the other major providers, which makes switching from one to another easier:
export MISTRAL_API_KEY="your-key"
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":"Explain NAT in three sentences."}]}'
As with any API key: never in a Git repository or in a web page's code, always in a secrets manager or an environment variable.
5. Running Mistral yourself
The open models can be downloaded from Hugging Face and used with Ollama, LM Studio or vLLM. For a homelab, the Ministral and Mistral Small models are the best suited:
ollama pull ministral-3
ollama run ministral-3 \
"Write an nftables rule that only allows SSH from 10.10.10.0/24."
The exact model name in Ollama varies between versions: check it in the Ollama library. To set up a complete interface, see Ollama and Open WebUI in a homelab.
| Model | Indicative hardware |
|---|---|
| Ministral 3B | Recent laptop, no dedicated graphics card |
| Ministral 8B to 14B | 16 to 24 GB of memory (graphics card or Apple silicon Mac) |
| Mistral Small 4 (119B) | Server with several graphics cards, or a compressed (quantised) version on a very well-equipped machine |
The benefit is twofold: no data leaves your network, and the model stays available even if the online offer changes.
6. Privacy and training
| Plan | Are your conversations used for training? |
|---|---|
| Vibe Free and Pro (individuals) | Yes by default, can be turned off in the settings |
| Vibe Team | The administrator can turn off training for the whole organisation |
| Vibe Enterprise | No by default, managed by administrators |
| API (Mistral Studio) | "Anonymous improvement data" setting, to turn off in the Privacy menu of the admin area |
To turn off training:
- Web: admin.mistral.ai › Manage › Vibe › Privacy, turn off the training option.
- Mobile: profile › Settings › Account, Data & Account Controls section, untick Enable data sharing.
- The Vibe and API settings are independent: turn off both if you use both.
Mistral is a company under French law, subject to the GDPR. For professional use, check in the contract the hosting location, the retention period and the zero data retention options offered for enterprise plans.
7. Mistral and sovereignty
Mistral is often cited in digital sovereignty initiatives. What it really brings:
| Strength | What to check |
|---|---|
| European publisher, not directly subject to American extraterritorial laws | Where your data is hosted depending on the plan, and by whom (Mistral, a third-party cloud, your own infrastructure) |
| Self-hostable open models: maximum sovereignty | The computing power required and the skills to run it |
| Its own data centres in France | The certifications your sector requires for the chosen offer (for example SecNumCloud or HDS in France) |
| Private deployments for businesses | The contract, reversibility, support |
See the cloud comparison and On-premises or SaaS for the evaluation method.
8. Mistral, Gemini or Claude?
| Need | Natural choice |
|---|---|
| Sovereignty, data in Europe, self-hostable models | Mistral |
| Google ecosystem (Gmail, Drive, Android), images and video | Gemini (see the Google Gemini ecosystem) |
| Long writing, reasoning, tasks handed over end to end | Claude (see Getting started with Claude) |
| Coding with agents | All three: Vibe CLI, Antigravity, Claude Code |
| Homelab, no data leaving | Open models: Ministral, Mistral Small, or Google's Gemma |
Key points
- Le Chat has been called Vibe since May 2026: a Work mode for office tasks, a Code mode for development (CLI, IDE, web).
- Mistral publishes many open models under Apache 2.0, usable at home with Ollama.
- Mistral Large 4 (October 2026) is its largest model; its weights are due to be published after safety testing.
- On the Free and Pro plans, training on your conversations is on by default: turn it off in the admin area and on mobile.
- Mistral is a sovereignty asset, provided you check where the chosen offer is actually hosted.
Sources
- Mistral AI: Le Chat becomes Vibe (28 May 2026)
- Mistral AI: plans and features
- Mistral AI: Mistral Large 4 documentation
- TechCrunch: Mistral Large 4, a model of around 1 trillion parameters (6 October 2026)
- Mistral AI: list of models, licences and retired models
- Mistral AI: opting out of the use of your data for training
- W.Media (data centre trade press): Mistral's data centres (May 2026)