Mistral AI Models: Licenses, Web Search and Training Bot
Table of Contents
Mistral AI is a French company that makes large language models (LLMs). Some models have open weights (model files that anyone can download), and others are available only through the company API. A model does not browse the web by itself. For site owners, two parts matter: the MistralAI-Training crawler, which collects training data, and the web search tool that developers can give to an agent.
This page is about the models. For the assistant, see the Le Chat (now Vibe) guide.
Current models
The models page in the documentation shows a license or an access label for each model. These are the main general models in October 2026:
| Model | License or access |
|---|---|
| Mistral Large 4 | Open weight. Announced on October 6, 2026. |
| Mistral Large 3 | Apache 2.0 |
| Mistral Medium 3.5 | Modified MIT |
| Mistral Small 4 | Apache 2.0 |
| Ministral 3 (14B, 8B, 3B) | Apache 2.0 |
For Large 4, the company said that it will release the weights by the end of October 2026. Until then, the model is available through a preview API (announcement). Some other models, such as Codestral, have the label “Premier”, which is an API access tier.
Anyone can download an open-weight model and run it on their own servers. A robots.txt rule for MistralAI-Training applies only to that crawler. It does not cover apps that other companies build on downloaded models.
Web access through the API
A model generates text from the input that it gets. It does not search the web.
Developers can give an agent web access with the built-in websearch tool. The tool has two versions:
web_searchgives access to a search engine.web_search_premiumalso gives access to news articles from verified news providers.
The answer includes reference chunks with the title and URL of each source. The tool works with the Agents API and the Conversations API, not with the Chat Completions API. The documentation does not say which user agent this tool uses when it reads pages.
Training crawler
The company collects web content for model training with MistralAI-Training (crawler documentation). Its user agent string is:
Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; MistralAI-Training/1.0; +https://docs.mistral.ai/robots)The documentation says that site owners can disallow this user agent in robots.txt. If you want to keep your pages out of the training data, add this group:
User-agent: MistralAI-Training
Disallow: /To keep only one folder out of training, disallow that path:
User-agent: MistralAI-Training
Disallow: /research/This crawler is not used for search indexing or for live answers in the assistant. So the block does not change whether the assistant finds or cites your pages. No IP list is published for this crawler, so you cannot verify its requests by address. The crawler documentation also lists the two agents that serve the assistant, which the Le Chat guide covers.