Mistral AI
Conflict of interest. This site uses Claude, by Anthropic, to read and summarise texts; Anthropic is one of the companies on this page. Claude gives no judgements or rankings. The categories of the Seoul safety frameworks (e.g. «does not cover severe risks») come from Claude's reading of the documents: the document is linked on each company's page.
1 · Who it is
- Headquarters
- France · Europe
- Why it is on this page
- signed the Seoul commitments (May 2024)
- is rated by the FLI AI Safety Index
- Latest releases
- 29 models released in the last 24 months
- Mistral Large 4 6 Oct 2026
- Mistral Medium 3.5 29 Apr 2026
- Mistral Small 4 16 Mar 2026
- Devstral 2 (123B) 9 Dec 2025
- Devstral 2 (24B) 9 Dec 2025
2 · What it promised
Public commitments and signature status. Official lists have no dates: for names that appeared after our first reading of the list, the day we first saw them.
- signed
- signed
- not applicableFrozen archive: lists read from copies and articles on 26 October 2026 (the archive did not respond), to be rechecked.
- not on the list as of 27 Oct 2026List of submitted reports read on oecd.ai, without automated collection; report texts are not republished (all rights reserved).
- not applicableList read on AGORA (ETO/CSET, CC BY-NC, no automated collection) and copied into the file. The WeChat original cannot be opened by programs.
- EU AI Pact (opens in a new tab)25 Sep 2024not on the list as of 27 Oct 2026
3 · What it wrote and changed
Under the Seoul commitments the companies promised to publish, by the Paris summit (10-11 February 2025), a safety framework on severe risks, with the thresholds beyond which they consider them intolerable.
no safety framework found on the company's website as of 27 Oct 2026
Search by Claude of the company's websites on 27 Oct 2026, still to be checked again by a person.
Documents tracked over time
Versions of the text found in the copies of the Wayback Machine and Open Terms Archive: each point is a version that differs from the previous one.
Usage policy
Usage Policy (opens in a new tab)
Open Terms Archive tracks it as the «Vibe» service («Le Chat» until 9 July 2026).
Choose a point to read what changed.
First version seen on 10 Jul 2026
This version: Open Terms Archive (opens in a new tab)
Training data summary
AI Governance: Models (opens in a new tab)
List of Mistral AI models with their AI Act documentation, including the training data summaries.
Choose a point to read what changed.
First version seen on 5 Dec 2025
This version: archived copy (Wayback Machine) (opens in a new tab)
Terms of service
Terms of Service (EU consumers) (opens in a new tab)
Open Terms Archive tracks it as the «Vibe» service («Le Chat» until 9 July 2026).
Choose a point to read what changed.
First version seen on 10 Jul 2026
This version: Open Terms Archive (opens in a new tab)
Privacy policy
Privacy Policy (opens in a new tab)
Open Terms Archive tracks it as the «Vibe» service («Le Chat» until 9 July 2026).
Choose a point to read what changed.
First version seen on 10 Jul 2026
This version: Open Terms Archive (opens in a new tab)
Summaries by Claude, checked by a person; the text of the documents belongs to the company: read it through the links.
4 · What others say
Stanford Foundation Model Transparency Index
18 out of 100 (average of the 13 rated companies: 41) · Dec 2025 edition
Data: Stanford CRFM, CC BY 4.0 licence · data and method (opens in a new tab)
Ratings by others: see FLI AI Safety Index (Summer 2026) (opens in a new tab). Link only: their scores cannot be republished without permission.
Each rating measures different things on different companies: they are not comparable with each other.
5 · What happened
Incidents in which the company is named as developer of the system (AI Incident Database and AIAAIC).
- last 6 months
- 0
- last 12 months
- 3
- last 24 months
- 7
Severity, last 24 months: serious 0 · moderate 5 · minor 0 · not rated 2
- AI chatbots lure vulnerable gamblers to unlicensed betting websites (opens the Austria page)
- Reported Public Exposure of Over 100,000 LLM Conversations via Share Links Indexed by Search Engines and Archived
- Top chatbots tricked into generating instructions on how to enrich uranium
- Political chatbots provide biased political advice about Dutch elections (opens the Netherlands page)
- Top AI models generate misleading US election information 30 percent of the time (opens the United States page)
- Study: AI chatbots fail disabled voters (opens the United States page)
- Study: Language models gather and pass personal info to hackers
All incidents mentioning Mistral AI, in Explore incidents →Where the harm happens, in Connections →
6 · What we cannot see
- internal documents and unpublished safety evaluations;
- versions of documents that no archive saved;
- “no document found” is the result of a search, not proof that the document does not exist;
- signatory lists often have no dates: the date is the day we first saw the name;
- incidents the press does not report: the best-known companies have more recorded incidents because they are talked about more;
- many incidents do not have a severity yet;
- Chinese companies publish less in English and some of their commitments are read from third-party sources (AGORA).