LICENSEWARE
Contribute

Mistral AI licensing

From the LICENSEWARE WikiLast edited 19 ReferencesCited

This article is about the licenses for Mistral AI model weights (Apache 2.0, Modified MIT, Mistral Research License, Mistral Non-Production License, Mistral Commercial License) and Mistral's API and Vibe (formerly Le Chat) commercial terms. It is not legal advice.

On This Page

Mistral AI licensing covers two separate things: the licenses under which the Paris-based company Mistral AI releases model weights, and the commercial terms for its hosted services. Mistral releases open weights under four licenses: the Apache License 2.0, a modified MIT license, the Mistral AI Research License (MRL) and the Mistral AI Non-Production License (MNPL). Commercial self-deployment of MRL or MNPL models requires a separate Mistral Commercial License.[1][5][6][7] Hosted access is sold in two ways. The API is priced per million input and output tokens, with other meters for OCR, audio and tools. Vibe (formerly Le Chat) is sold as Free, Pro, Team and Enterprise plans, with Team priced per user per month.[16][17][19] The Commercial Terms of Service and the Additional Product Terms, both effective 2026-09-25, govern commercial customers.[14][15] For the general concepts, see Software license and License compliance.

License families and versions

Mistral Docs labels each model with its license: Apache 2.0, Modified MIT, CC BY-NC 4.0 (Voxtral TTS), or “Premier” (for example Codestral, OCR 4.1 and Mistral Embed).[2]

License family Catalog row Models named by Mistral Commercial use of self-hosted weights 
Apache License 2.0 Apache 2.0 open-weight models Mistral Large 3; Ministral 3 14B, 8B, 3B (2025-12-02); Mistral Small 4; Voxtral Small; Voxtral Mini Transcribe Realtime; Shieldstral 1.0; Leanstral 1.5 Permitted under Apache 2.0 terms[1][3][2] 
Modified MIT Modified MIT license models Mistral Medium 3.5 (released 2026-04-28) Permitted, but companies with monthly revenue above USD 20M need a commercial license or must use Mistral Studio[1][4] 
Mistral AI Research License MRL Mistral Large 2.0 (v24.07, retired) Research Purposes only; needs a Mistral Commercial License[5][7][9] 
Mistral AI Non-Production License MNPL Codestral (v24.05) Non-production use only; needs a license from Mistral AI[6][8][10] 

The MRL and MNPL texts carry version numbers (0.1) but state no effective date.[5][6] The catalog therefore dates MNPL rules to its introduction with Codestral on 2024-05-29, and MRL rules to the release of Mistral Large 2 under that license on 2024-07-24.[8][7] Both licenses are governed by French law, with the courts of Paris having exclusive jurisdiction.[5]

The Apache License 2.0 is published by the Apache Software Foundation and appears on the Open Source Initiative’s list of approved licenses.[11][12] The modified MIT text is not published on the Mistral-owned pages cited here. The Help Center says the licensing terms are documented in the model cards.[1]

Conditions and thresholds

  • Monthly revenue (Modified MIT). Companies with monthly revenue above USD 20M must either obtain a commercial license by contacting Mistral AI or use the models via Mistral Studio (rule). Mistral does not publish how revenue is measured.[1]
  • Research Purposes (MRL). Models, Derivatives and Outputs may be used only for Research Purposes: personal, scientific or academic research that is non-profit and non-commercial (rule). The definition excludes use by company employees or contractors in their daily tasks, any revenue-oriented activity including testing or proof-of-concept, and any distribution by a commercial entity, including through a hosted or managed service.[5]
  • Non-Production Environments (MNPL). Use is limited to testing, research, Personal or evaluation purposes in Non-Production Environments. The models may not be supplied in the course of a commercial activity, including as SaaS or behind a software layer (rule). A Non-Production Environment excludes live conditions, commercial operations, revenue-generating activities and direct impacts on end users.[6]
  • Other uses. Under both licenses, any use they do not authorise requires a license from Mistral AI, “which Mistral AI may grant to You in Mistral AI’s sole discretion” (rule).[5][6] For Mistral Large 2, Mistral said that commercial use requiring self-deployment needs a Mistral Commercial License. Its terms are not published.[7]

The pricing FAQ describes self-hosting differently from the Help Center. It says that open-weight models such as Mistral 7B are “Apache 2.0 licensed for research/individual use”, and that commercial deployments require a Mistral license.[16] The Help Center says Apache 2.0 allows use of the models for any purpose.[1] The license text itself is the Apache License 2.0.[11]

Acceptable use

The Mistral AI Usage Policy, effective 2026-06-11, applies to products on Mistral’s own platform, including Vibe and Mistral AI Studio. It prohibits illegal activity, child sexual abuse material, fraud and scams, and privacy violations. It states that it does not apply to products deployed on a customer’s or partner’s infrastructure, or to Mistral’s open-source models (rule).[13]

The Commercial Terms of Service (Section 2.2) forbid several things (rule):[14]

  • reverse engineering the products or their weights, or using Output to do so;
  • buying, selling or transferring API keys or accounts;
  • integrating Vibe into the customer’s own products offered to third parties without written authorisation.

The customer owns Output, and model weights and parameters are not Output (rule).[14]

Redistribution and derivatives

  • Apache 2.0. Redistribution requires giving recipients a copy of the license, marking modified files, keeping copyright and attribution notices, and passing on any NOTICE file contents.[11]
  • MRL and MNPL. Distributors must make the agreement available to recipients, whose rights come directly from Mistral AI. They must also keep a “Notice” file reading “Licensed by Mistral AI under the Mistral AI Research License” or ”… Non-Production License” (MRL rule, MNPL rule).[5][6]
  • Your own derivatives under the MRL. These may be distributed under different terms, provided they state prominently that the model was modified and the terms do not conflict with the MRL. Any Derivative remains bound by the license’s usage limitation.[5]
  • Naming. Neither license requires a name prefix. A licensee must not imply that a Derivative is an official Mistral AI product or endorsed by Mistral AI, and no trademark license is granted beyond descriptive use and the required attribution (rule).[5][6]

Commercial API and plans

API

Under the Commercial Terms, pricing is whatever the pricing page shows, unless an Order Form says otherwise. Customer Accounts are charged monthly, and paid fees are non-refundable (rule).[14] Mistral’s 2024 announcements called the API platform “la Plateforme”; current pricing pages refer to Mistral Studio.[7][16]

The pricing FAQ gives the method (rule):[16][17]

  • input and output tokens are counted separately, per million tokens;
  • batch processing costs 50% less;
  • cached input costs up to 90% less;
  • OCR is priced per 1,000 pages, speech per minute, and tool APIs per call;
  • regional inference endpoints add 10% on supported models.

Standard-tier list prices in USD per million tokens, as of 2026-09-26:[17][18]

Model License label Input Cached input Output 
Mistral Large 3 Apache 2.0 0.5 0.05 1.5 
Mistral Medium 3.5 Modified MIT 1.5 0.15 7.5 
Mistral Small 4 Apache 2.0 0.15 0.015 0.6 
Ministral 3 14B Apache 2.0 0.2 0.02 0.2 
Ministral 3 8B Apache 2.0 0.15 0.015 0.15 
Ministral 3 3B Apache 2.0 0.1 0.01 0.1 
Codestral Premier 0.3 not listed 0.9 

Other list prices:[17]

  • OCR 4.1: USD 4 per 1,000 pages for OCR, USD 5 for Document AI.
  • Voxtral Mini Transcribe 2: USD 0.003 per minute of audio.
  • Voxtral TTS: USD 0.016 per 1,000 characters.
  • Web search and code execution: USD 30 per 1,000 calls each.

Enterprise APIs, which add regional data processing controls, SLAs, higher rate limits and premium support, cost 75% above list on select APIs (rule).[17]

Vibe plans

Le Chat was renamed Vibe, and existing accounts and plans carried over.[19] As of 2026-09-26 the plans page listed:[16]

Plan Catalog row Price (excl. taxes) Notes 
Free Vibe Free USD 0 Limited messages; USD 10 / month in API credits 
Pro Vibe Pro USD 14.99 / month Students USD 5.99 / month for up to 12 months; fair usage limits 
Team Vibe Team USD 24.99 / user / month Up to 30GB storage per user; domain verification; data export 
Enterprise Vibe Enterprise On request Custom models; audit logs; SAML SSO; white label; self-hosted, private cloud or on-premises deployment 

Seats are End User Accounts created under the Customer Account, up to any maximum number of seat licenses. They may not be shared outside the customer’s organization (rule).[15] A customer that chooses a fixed-term annual or multi-annual agreement instead of a monthly one owes the remaining fees as an Early Termination Fee if it terminates early (rule).[15] For how these meters compare with other vendors, see Subscription and consumption licensing and Meta Llama licensing.

Out of scope

  • Third-party models served through Mistral’s API, such as Z.ai GLM 5.x. These are governed by their own licenses.
  • Models licensed under CC BY-NC 4.0 (Voxtral TTS). The license label is recorded here, but this page does not analyse the license.
  • The price of the Mistral Commercial License, and the modified MIT license text. Neither is published on the Mistral pages cited here.
  • Consumer terms of service, the Data Processing Addendum, partner-served deployment terms, and Forge and AI Cloud pricing.

Catalog rows

Catalog rows for this article, grouped by table.

References

  1. Under which license are Mistral’s open models available?Mistral Help Center, updated 12 August 2026. Apache 2.0; modified MIT with USD 20M monthly revenue condition.Effective 2026-08-12. Retrieved 2026-09-26.
  2. Models Overview (Mistral Docs)License label per model; deprecation and retirement table.Retrieved 2026-09-26.
  3. Introducing Mistral 3Official announcement, 2 December 2025.Effective 2025-12-02. Retrieved 2026-09-26.
  4. Mistral Medium 3.5 (Mistral Docs model page)Released 28 April 2026; Modified MIT; USD 1.5 / 7.5 per M tokens.Effective 2026-04-28. Retrieved 2026-09-26.
  5. Mistral AI Research License (MRL-0.1)License text; no effective date stated.Retrieved 2026-09-26.
  6. Mistral AI Non-Production License (MNPL-0.1)License text; no effective date stated.Retrieved 2026-09-26.
  7. Large Enough (Mistral Large 2 announcement)Official announcement, 24 July 2024. MRL; Mistral Commercial License for self-deployment.Effective 2024-07-24. Retrieved 2026-09-26.
  8. The Mistral AI Non-Production LicenseOfficial announcement, 29 May 2024.Effective 2024-05-29. Retrieved 2026-09-26.
  9. Mistral Large 2.0 (Mistral Docs model page)Open weights under MRL; retired.Retrieved 2026-09-26.
  10. Codestral 24.05 (Mistral Docs model page)Released under the MNPL.Retrieved 2026-09-26.
  11. Apache License, Version 2.0Apache Software Foundation, January 2004.Retrieved 2026-09-26.
  12. Apache License, Version 2.0 (Open Source Initiative)OSI license list entry.Retrieved 2026-09-26.
  13. Mistral AI Usage PolicyEffective 11 June 2026.Effective 2026-06-11. Retrieved 2026-09-26.
  14. Mistral AI Commercial Terms of ServiceEffective 25 September 2026.Effective 2026-09-25. Retrieved 2026-09-26.
  15. Mistral AI Additional Product TermsEffective 25 September 2026.Effective 2026-09-25. Retrieved 2026-09-26.
  16. Mistral pricing (plans)Vibe Free, Pro, Team, Enterprise; pricing FAQ. Undated.Retrieved 2026-09-26.
  17. Mistral API pricingPer-model API prices; Enterprise APIs; regional inference. Undated.Retrieved 2026-09-26.
  18. Pricing (Mistral Docs)Input, cached input and output per M tokens. Undated.Retrieved 2026-09-26.
  19. Le Chat is now VibeMistral Help Center, 12 August 2026.Effective 2026-08-12. Retrieved 2026-09-26.

See also

Esc