# Mistral AI

> Source: https://aiwiki.ai/wiki/mistral_ai
> Updated: 2026-07-28
> Fact-checked: 2026-07-28
> Categories: AI Companies, Artificial Intelligence, Large Language Models, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Mistral AI." aiwiki.ai, 28 Jul 2026. https://aiwiki.ai/wiki/mistral_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Mistral AI is a French [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company that develops [large language models](https://aiwiki.ai/wiki/large_language_model), multimodal models, software for building and operating AI systems, and computing infrastructure. Its legal entity, Mistral AI SAS, is registered in Paris under SIREN 952 418 325, and Arthur Mensch is its chief executive officer.[1] France's SIRENE register records the enterprise as active from April 28, 2023.[2]

The company was founded by Mensch, Guillaume Lample, and Timothee Lacroix. As of July 2026, Mistral identified Mensch as CEO, Lample as chief science officer, and Lacroix as chief technology officer.[3] Mistral first became known for releasing downloadable model weights, then expanded into commercial APIs, an assistant, enterprise software, model customization, and a private compute offering. Its releases have used several different licenses, so the company and its models should not be described with a single blanket label such as "open source."

## Corporate identity and founding

Mensch had worked at Google DeepMind, while Lample and Lacroix had worked at Meta before forming the company. Mistral announced a EUR 105 million seed round in June 2023, about six weeks after the company's registered start date. Lightspeed Venture Partners led the round. Contemporary reporting placed the post-money valuation at approximately EUR 240 million, but that valuation was reported rather than disclosed in a public filing.[4]

Mistral positioned itself from the beginning as a Europe-based developer of [generative AI](https://aiwiki.ai/wiki/generative_ai) systems for organizations that wanted more deployment control than an API-only service provided. That position later developed into three connected lines of business:

- models distributed as downloadable weights or through hosted APIs;
- applications and enterprise development products;
- infrastructure and services for private or regionally controlled deployment.

This combination is important to the company's identity. Mistral is not only a model publisher, and its consumer assistant is not the whole company. It sells access to hosted models, software for evaluating and operating AI applications, custom model work, and compute capacity. It also supports self-deployment for models whose licenses and technical requirements allow it.

## Financing and governance

Mistral's early financing occurred in unusually rapid succession. In December 2023, it closed a EUR 385 million Series A led by Andreessen Horowitz, with Lightspeed participating. Reporting at the time placed the company near a USD 2 billion valuation.[5] In June 2024, a Series B raised approximately EUR 600 million in a mix of equity and debt. General Catalyst led the round, and contemporary reporting placed the valuation at about USD 6 billion.[6]

In September 2025, Mistral announced a EUR 1.7 billion Series C at a EUR 11.7 billion post-money valuation. [ASML](https://aiwiki.ai/wiki/asml), the Dutch semiconductor-equipment manufacturer, led the round, with existing investors including Andreessen Horowitz, Bpifrance, General Catalyst, Lightspeed, and [Nvidia](https://aiwiki.ai/wiki/nvidia) also participating.[7] ASML separately disclosed that it invested EUR 1.3 billion, expected to hold about 11 percent of Mistral on a fully diluted basis, and received a seat on Mistral's Strategic Committee. ASML described that seat as advisory and linked the investment to joint work on AI-enabled products and research.[8]

The company's 2024 relationship with [Microsoft](https://aiwiki.ai/wiki/microsoft) drew regulatory attention because Microsoft both supplied cloud infrastructure and invested in the company. The United Kingdom Competition and Markets Authority found that Microsoft's investment consisted of EUR 15 million in convertible bonds, with any resulting equity interest expected to be below 1 percent. The authority also found that Microsoft had no board representation or other material governance rights, and that the Azure distribution arrangement was nonexclusive. It concluded that Microsoft did not acquire material influence over Mistral, so the arrangement did not qualify as a relevant merger situation under UK law.[9] This was a jurisdictional finding, not a finding that Microsoft controlled Mistral or that every aspect of the partnership had been reviewed as a completed merger.

Published financing totals require care. Mistral's rounds have included equity, convertible bonds, conventional debt, and later project financing. Adding every announced amount as if it were common equity would overstate the capital invested on that basis. The company is privately held, and public sources do not provide a complete current capitalization table.

## Model development

Mistral's first public model, [Mistral 7B](https://aiwiki.ai/wiki/mistral_7b), was released in September 2023.[3] Its technical report described a 7-billion-parameter decoder model using [grouped-query attention](https://aiwiki.ai/wiki/grouped_query_attention) and [sliding window attention](https://aiwiki.ai/wiki/sliding_window_attention). The weights were released under the Apache License 2.0.[10] The report compared the model with larger systems on selected benchmarks, but those author-reported 2023 results should not be read as a current ranking.

The company next released [Mixtral](https://aiwiki.ai/wiki/mixtral), a sparse [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) model. In each feed-forward layer, a router selected two of eight experts for each token. The technical report said the model had 47 billion total parameters but used 13 billion active parameters for a token, and released both base and instruction-tuned versions under Apache 2.0.[11] This established sparse expert routing as a recurring part of Mistral's model strategy.

Mistral later moved beyond text-only systems. The October 2024 [Pixtral](https://aiwiki.ai/wiki/pixtral) report described a 12-billion-parameter multimodal model with a vision encoder trained from scratch, support for variable image sizes, and an Apache 2.0 release.[12] [Magistral](https://aiwiki.ai/wiki/magistral), introduced in June 2025, was the company's first model family explicitly trained for extended reasoning. Its report described a reinforcement-learning pipeline and distinguished the commercial Magistral Medium from the Apache-licensed Magistral Small.[13]

In December 2025, the Mistral 3 release grouped [Mistral Large 3](https://aiwiki.ai/wiki/mistral_large_3) with new [Ministral](https://aiwiki.ai/wiki/ministral) models. Mistral released base and instruction-tuned versions under Apache 2.0 and described the family as multimodal.[14] Other company model lines have focused on code, including [Codestral](https://aiwiki.ai/wiki/codestral), and on audio, including [Voxtral](https://aiwiki.ai/wiki/voxtral). Detailed architecture, benchmark, and version histories belong on those model pages because availability and specifications change more quickly than company-level facts.

## Licensing and the meaning of openness

Mistral uses a mixed licensing strategy. Some model weights and associated code have been released under Apache 2.0, which permits commercial use subject to the license terms. Other releases have used research or commercial terms. In May 2024, for example, Mistral introduced the Mistral AI Non-Production License for Codestral, allowing research and noncommercial use while requiring a separate license for production use.[15]

The distinction between [open weights](https://aiwiki.ai/wiki/open_weights) and [open-source AI](https://aiwiki.ai/wiki/open_source_ai) is material. The Open Source Initiative's Open Source AI Definition requires freedoms to use, study, modify, and share the system, together with access to the preferred form for modification. For machine-learning systems, that includes model parameters, code, and sufficiently detailed information about the training data.[16] Mistral states that it does not disclose the datasets used to train its models and that its training datasets, training logic, and training resources are proprietary.[17] Under the OSI definition, releasing weights under a permissive license therefore does not by itself establish that the complete trained system is open-source AI.

This does not make all Mistral releases closed in the same way. A user may be able to download, inspect, fine-tune, and self-host Apache-licensed weights even though the original training process cannot be reproduced from public materials. Conversely, a hosted commercial model may expose only an API. Descriptions of a particular release should name its actual license and disclosed artifacts rather than infer them from Mistral's general branding.

## Distribution and enterprise products

Mistral opened a hosted developer platform, commonly called La Plateforme and later incorporated into Mistral AI Studio, alongside its downloadable models. In February 2024, Microsoft announced a multi-year partnership that provided Mistral with access to [Microsoft Azure](https://aiwiki.ai/wiki/azure) infrastructure and distributed Mistral models through Azure's model catalogue.[18] Mistral's own deployment documentation later listed commercial model availability through Azure AI, [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services) Bedrock, [Google Cloud](https://aiwiki.ai/wiki/google_cloud) Vertex AI, Snowflake Cortex, IBM watsonx, and Outscale.[19] Provider and model availability varies by region and version, so that list is a distribution record rather than a guarantee that every Mistral model is present on every service.

The company's product layer broadened after its initial API and assistant:

- [Le Chat](https://aiwiki.ai/wiki/le_chat) began as a conversational interface to Mistral models. By the research cutoff, Mistral's current product navigation presented [Mistral Vibe](https://aiwiki.ai/wiki/mistral_vibe) for work and coding while retaining Le Chat in its company timeline.[3]
- Mistral AI Studio, announced in October 2025, added facilities for versioning prompts, running evaluations, monitoring applications, collecting feedback, fine-tuning, and operating governed AI workflows.[20]
- [Mistral Forge](https://aiwiki.ai/wiki/mistral_forge), announced in March 2026, is a system for pretraining, post-training, reinforcement learning, and evaluating models using an organization's proprietary data and requirements. Mistral named ASML, Ericsson, the European Space Agency, and Singapore public-sector organizations among early users.[21]

These products support a business model that combines usage-based hosted services, software subscriptions, licensing, and enterprise implementation work. The model catalogue is an input to that business, but enterprise customers may buy a deployment, customization process, or application platform rather than a named model alone.

## Compute infrastructure

Mistral announced Mistral Compute in June 2025 as a private, integrated infrastructure offering spanning GPUs, orchestration, APIs, products, and services. The company said it would offer configurations from bare-metal systems to a managed platform and would continue making its models available through public clouds and on-premises deployments.[22] The offering placed Mistral in a more capital-intensive part of the AI supply chain than model development alone.

In March 2026, the law firm Gide disclosed financing of approximately USD 830 million for Mistral's data-center project at Bruyeres-le-Chatel near Paris. The package consisted of a roughly USD 720 million tranche and a EUR 94 million tranche from a seven-bank syndicate. Gide said the financing would support the acquisition of 13,800 Nvidia GB300 GPUs and development of the site's compute capacity.[23] These are financed project plans and procurement figures, not proof that every GPU was installed or operational on the announcement date.

At its May 2026 AI Now event, Mistral also announced a 10-megawatt inference facility at Les Ulis and said it was scheduled to open in the third quarter of 2026.[24] Because that date was still in the future at the July 28, 2026 research cutoff, the facility should be described as announced or scheduled unless later operational evidence is available.

Owning compute can give Mistral more control over capacity, location, and deployment terms, but it also changes its financial exposure. Data centers require large up-front commitments and ongoing utilization. Mistral has not published enough audited project-level information to calculate the utilization, profitability, or full environmental footprint of its planned infrastructure.

## Acquisition and industrial strategy

In May 2026, Mistral announced a definitive agreement to acquire Emmi AI, an Austrian company developing machine-learning systems for engineering and physical simulation. The announcement said more than 30 Emmi researchers and engineers would join Mistral, but did not disclose transaction terms.[25] Mistral subsequently described physics-focused models and industrial engineering tools as a strategic expansion beyond general-purpose language applications.

The same AI Now event identified Airbus, BMW, and ASML as partners for industrial engineering work.[24] These relationships illustrate Mistral's emphasis on manufacturing, semiconductors, transport, energy, finance, and government. They do not independently establish cost savings or performance gains for customers. Claims about deployment outcomes should be tied to a named customer record, a stated measurement period, and a disclosed method.

The ASML relationship is broader than a financing transaction. ASML linked its investment to the application of Mistral models across its product portfolio and to possible joint research.[8] This makes industrial customization and proprietary-data work a significant part of Mistral's strategy, alongside general-purpose model releases.

## Regulation, safety, and transparency

Mistral participates in both voluntary and statutory governance processes. It was one of the companies that signed the Frontier AI Safety Commitments at the 2024 [AI Seoul Summit](https://aiwiki.ai/wiki/seoul_ai_summit). The signatories committed to publish severe-risk safety frameworks, define risk thresholds, evaluate systems, apply mitigations, and report on their approach.[26] These were voluntary commitments and did not themselves certify that a signatory had implemented every measure.

Mistral also signed the European Union's General-Purpose AI Code of Practice. The European Commission describes the code as a voluntary tool for demonstrating compliance with obligations for providers of general-purpose AI models; its chapters cover transparency, copyright, and safety and security.[27] Mistral's Legal Center publishes model and system documentation related to the EU AI Act and other applicable rules.[28] The existence of such documentation is not equivalent to an independent finding of compliance for every model or deployment.

An external review has identified a gap between Mistral's commitments and public company-level safety governance. The Future of Life Institute's Summer 2026 AI Safety Index reported that Mistral had not published its own safety framework and gave the company the lowest overall score among the nine developers it assessed.[29] That result is an evaluation by an advocacy organization using its own methodology, not a regulatory judgment. Still, the specific framework finding is relevant because the Seoul commitments called for signatories to publish such a framework.

Mistral does publish [model cards](https://aiwiki.ai/wiki/model_card), acceptable-use rules, moderation tools, and regulatory documentation for particular products. Those artifacts answer narrower questions about a model or service. A frontier safety framework would instead state company-level capability thresholds, escalation rules, evaluation procedures, and conditions under which development or deployment would be paused. The two forms of documentation should not be treated as interchangeable.

## Environmental reporting

In July 2025, Mistral published a life-cycle assessment of Mistral Large 2 developed with consultancy Carbone 4 and the French ecological-transition agency ADEME, with review by two environmental consultancies. The assessment covered greenhouse-gas emissions, water consumption, and resource depletion across model development and use.[30] Mistral described the work as a first approximation and noted important methodological limits, including the absence of standardized public impact factors for AI hardware.

ADEME separately described the study as an effort to develop comparable indicators and said its scope covered the model's life cycle through January 2025.[31] The assessment is therefore evidence that Mistral disclosed a bounded estimate for one model under a stated method. It should not be generalized into an audited footprint for all later models, cloud partners, or data-center projects.

## Limitations and unresolved questions

Several kinds of claims about Mistral remain difficult to verify from public evidence:

- The company is private and does not publish audited revenue, profit, cash-flow, or current headcount statements on a regular schedule. Press estimates should be dated and attributed rather than presented as settled facts.
- Training-corpus composition is not public. This limits independent analysis of copyright, provenance, representativeness, contamination, and reproducibility.[17]
- Company benchmark charts generally report results selected and run by the developer. They can document what Mistral measured but do not by themselves establish broad superiority across tasks, languages, deployment settings, or later competing models.
- Open-weight access improves inspectability and deployment choice but does not reveal the original data pipeline or training process.
- Announced compute capacity and future facilities should not be counted as operational capacity until their commissioning is independently documented.

Data practices also differ by product tier. As of the research cutoff, Mistral's help center said that inputs and outputs from Free, Pro, and Education products could be used for model training by default unless the user opted out, while Team and Enterprise data was not used for that purpose. Studio Free and Studio Scale had separate controls.[32] These terms are mutable and should be checked at the time of use rather than treated as permanent company policy.

Mistral's significance comes from the combination of a European corporate base, a mixed open-weight and commercial model strategy, broad cloud distribution, enterprise customization, and investment in its own compute. Each part has a different evidentiary standard. Model releases can be checked against papers and licenses; financing against company, investor, and regulatory records; and future infrastructure against commissioning evidence. Keeping those categories separate avoids turning company announcements into claims that projects, performance, or compliance have already been independently established.

## References

[1] Mistral AI, "Legal Notice," current legal identity and registered office. https://legal.mistral.ai/legal-notice
[2] Institut national de la statistique et des etudes economiques, "Avis de situation au repertoire Sirene: Mistral AI," accessed March 8, 2026. https://api-avis-situation-sirene.insee.fr/identification/pdf/95241832500017
[3] Mistral AI, "About Mistral," company history, founders, roles, and dated milestones. https://mistral.ai/about/
[4] Ingrid Lunden, "France's Mistral AI blows in with a $113M seed round," TechCrunch, June 13, 2023. https://techcrunch.com/2023/06/13/frances-mistral-ai-blows-in-with-a-113m-seed-round-at-a-260m-vaLuation-to-take-on-openai/
[5] Romain Dillet, "Mistral AI closes its $415 million funding round," TechCrunch, December 11, 2023. https://techcrunch.com/2023/12/11/mistral-ai-a-paris-based-openai-rival-closed-its-415-million-funding-round/
[6] Romain Dillet, "Paris-based AI startup Mistral AI raises $640M," TechCrunch, June 11, 2024. https://techcrunch.com/2024/06/11/paris-based-ai-startup-mistral-ai-raises-640-million/
[7] Mistral AI, "Mistral AI raises EUR 1.7B to accelerate technological progress with AI," September 9, 2025. https://mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai/
[8] ASML, "ASML, Mistral AI enter strategic partnership," September 9, 2025. https://www.asml.com/en/news/press-releases/2025/asml-mistral-ai-enter-strategic-partnership
[9] UK Competition and Markets Authority, "Microsoft Corporation's partnership with Mistral AI: full text decision," May 21, 2024. https://assets.publishing.service.gov.uk/media/664c6cfd993111924d9d389f/Full_text_decision.pdf
[10] Albert Q. Jiang et al., "Mistral 7B," arXiv:2310.06825, October 2023. https://arxiv.org/abs/2310.06825
[11] Albert Q. Jiang et al., "Mixtral of Experts," arXiv:2401.04088, January 2024. https://arxiv.org/abs/2401.04088
[12] Pravesh Agrawal et al., "Pixtral 12B," arXiv:2410.07073, October 2024. https://arxiv.org/abs/2410.07073
[13] Mistral AI et al., "Magistral," arXiv:2506.10910, June 2025. https://arxiv.org/abs/2506.10910
[14] Mistral AI, "Introducing Mistral 3," December 2, 2025. https://mistral.ai/news/mistral-3/
[15] Mistral AI, "The Mistral AI Non-Production License," May 29, 2024. https://mistral.ai/news/mistral-ai-non-production-license-mnpl/
[16] Open Source Initiative, "The Open Source AI Definition 1.0," October 2024. https://opensource.org/ai/open-source-ai-definition
[17] Mistral AI Help Center, "Does Mistral disclose its training datasets?" accessed July 28, 2026. https://help.mistral.ai/en/articles/347390-does-mistral-disclose-its-training-datasets
[18] Eric Boyd, "Microsoft and Mistral AI announce new partnership," Microsoft Azure Blog, February 26, 2024. https://azure.microsoft.com/en-us/blog/microsoft-and-mistral-ai-announce-new-partnership-to-accelerate-ai-innovation-and-introduce-mistral-large-first-on-azure/
[19] Mistral AI, "Cloud deployments," Mistral documentation, accessed July 28, 2026. https://docs.mistral.ai/models/deployment/cloud-deployments
[20] Mistral AI, "Introducing Mistral AI Studio," October 24, 2025. https://mistral.ai/news/ai-studio/
[21] Mistral AI, "Introducing Forge," March 17, 2026. https://mistral.ai/news/forge/
[22] Mistral AI, "Mistral Compute," June 11, 2025. https://mistral.ai/news/mistral-compute/
[23] Gide Loyrette Nouel, "Gide conseille Mistral AI dans le cadre de son premier financement," March 30, 2026. https://www.gide.com/news-insights/gide-conseille-mistral-ai-dans-le-cadre-de-son-premier-financement-dun-montant-de-c-830-millions-de-dollars-americains-afin-doperer-un-data-center-en-region-parisienne-a-bruyeres-le-chatel-franc/
[24] Mistral AI, "AI Now Summit 2026," May 28, 2026. https://mistral.ai/news/ai-now-summit-2026/
[25] Mistral AI, "Emmi joins Mistral to accelerate the AI-native industry," May 23, 2026. https://mistral.ai/fr/news/accelerate-ai-native-industry/
[26] UK Department for Science, Innovation and Technology, "Frontier AI Safety Commitments, AI Seoul Summit 2024," May 21, 2024. https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024
[27] European Commission, "The General-Purpose AI Code of Practice," updated April 23, 2026. https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
[28] Mistral AI, "Legal Center: AI Governance," accessed July 28, 2026. https://legal.mistral.ai/
[29] Future of Life Institute, "AI Safety Index: Summer 2026," July 2026. https://futureoflife.org/wp-content/uploads/2026/07/AI-Safety-Index-Summer-2026-Digital.pdf
[30] Mistral AI, "Our contribution to a global environmental standard for AI," July 22, 2025. https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai/
[31] Agence de la transition ecologique, "Nous avons besoin d'indicateurs fiables pour comparer les IA entre elles," 2025. https://infos.ademe.fr/societe-politiques-publiques/2025/nous-avons-besoin-dindicateurs-fiables-pour-comparer-les-ia-entre-elles/
[32] Mistral AI Help Center, "Do you use my user data to train your artificial intelligence models?" accessed July 28, 2026. https://help.mistral.ai/en/articles/347617-do-you-use-my-user-data-to-train-your-artificial-intelligence-models