CtrlK
| Mistral AI SAS | |
|---|---|
| |
| Type | Private |
| Industry | Artificial intelligence |
| Founded | April 28, 2023 |
| Founders | Arthur Mensch (CEO) Guillaume Lample (Chief Scientist) Timothée Lacroix (CTO) |
| Headquarters | Paris, France |
| Key people | Arthur Mensch (CEO) Guillaume Lample (Chief Scientist) Timothée Lacroix (CTO)
|
| Products | LLMs (open-weight and proprietary) Le Chat (assistant) La Plateforme (API platform) Mistral models Mixtral models Codestral Pixtral
|
| Valuation | €11.7 billion (Series C, Sept. 2025) |
| Employees | 350 (Sept. 2025) |
| Website | mistral.ai |
Mistral AI is a French artificial intelligence company headquartered in Paris, France. Founded in April 2023 by former researchers from Google DeepMind and Meta Platforms, Arthur Mensch, Guillaume Lample, and Timothée Lacroix, the company develops large language models (LLMs) and related tooling. The three co-founders originally met during their studies at École Polytechnique.[1] Mistral publishes both open-source/open-weight models (for example Mistral 7B, Mixtral 8x7B) and proprietary offerings delivered via API (for example Mistral Large, Mistral Medium 3, Magistral Medium).[1] In September 2025, it announced a €1.7 billion Series C led by ASML at an €11.7 billion post-money valuation, and reported a workforce of roughly 350 people across multiple offices.[2][3]
Mistral AI was incorporated on April 28, 2023, by Mensch, Lample, and Lacroix, alumni of École Polytechnique and veterans of Google DeepMind and Meta AI. Mensch spent nearly three years at Google DeepMind, while Lample and Lacroix are large-scale AI model specialists who had previously worked for Meta, with Lample being one of the creators of Meta's LLaMA language model.[1] The founders aimed to build AI models that are open, efficient, and capable of competing with those from larger U.S.-based companies.
In June 2023, one month after founding, the company raised a record €105 million seed round led by Lightspeed Venture Partners, including participation from investors such as Eric Schmidt, Xavier Niel, and JCDecaux Holding.[4][5]
In September 2023, Mistral released its first open-weight model, Mistral 7B, under the Apache License 2.0, including distribution via a BitTorrent magnet link and later on Hugging Face. This release was notable for being distributed via a torrent file, which sparked some controversy regarding distribution methods but highlighted the model's accessibility.[6][7]
In December 2023, Mistral announced Mixtral 8x7B, a sparse mixture of experts (SMoE) open-weight model.[8] That month it also raised €385 million in a second round led by Andreessen Horowitz (a16z) and Lightspeed, bringing the company's valuation to approximately $2 billion. The round included investment from Nvidia and Salesforce.[9]
In February 2024, Microsoft and Mistral announced a partnership to offer Mistral's models via Microsoft Azure, alongside a €15 million Microsoft investment structured to convert into equity in a subsequent round. The deal made Mistral's models available on the Azure cloud computing platform and established a research and development collaboration aimed at building applications for government and commercial customers. The arrangement drew review from EU competition authorities.[10][11][12]
In June 2024, Mistral raised €600 million at ~€5.8 billion valuation to scale compute and go-to-market.[13][14]
On September 9, 2025, the company announced a €1.7 billion Series C led by ASML, at €11.7 billion post-money; ASML became the largest shareholder (~11%).[15][2]
Mistral offers a family of text and multimodal LLMs accessible via API and through partner clouds (for example Azure AI, Amazon SageMaker, GitHub Models). Releases include open-weight and proprietary models geared toward cost-efficient inference, reasoning, coding assistance, embeddings, OCR, and audio input.[16][17][18]
| Date | Model | Type / modality | Parameters (approx.) | License / availability | Notes / source |
|---|---|---|---|---|---|
| 2023-09-27 | Mistral 7B | Text LLM | 7.3B | Apache-2.0 (open weights) | First release; distributed via torrent and Hugging Face. Utilizes Grouped-Query Attention (GQA) and Sliding Window Attention (SWA) for faster inference and reduced memory usage.[6][7] |
| 2023-12-11 | Mixtral 8x7B | Text LLM (SMoE) | ~46.7B (8×7B experts; ~12.9B active) | Apache-2.0 (open weights) | Open-weight mixture of experts model. For each token, it routes through two of its eight "expert" networks, significantly reducing computational cost while achieving the performance of a much larger dense model.[8] |
| 2024-04-10 | Mixtral 8x22B | Text LLM (SMoE) | 141B (total); 39B (active) | Apache-2.0 (open weights) | Larger MoE model with 64K token context window. Released via a torrent link on X.[19] |
| 2024-02-26 | Mistral Large | Text LLM (proprietary) | Unknown | Commercial API | The company's flagship model, designed for complex reasoning tasks. Fluent in multiple languages including French, German, Spanish, and Italian.[20] |
| 2024-05-29 | Codestral | Code LLM | ~22B (initial) | Open weights (permissive) | First code-specialist model. Supports 80+ programming languages including Python, Java, C++, JavaScript, and Bash.[21] |
| 2024-07 | Codestral Mamba | Code LLM | 7B | Apache-2.0 (open weights) | Uses the Mamba architecture for efficient code generation with linear time inference.[22] |
| 2024-09 | Pixtral 12B | Multimodal LLM | 12B + 400M vision encoder | Apache-2.0 (open weights) | First multimodal model capable of processing both text and images.[23] |
| 2024-11-18 | Le Chat (product) | Assistant UI (chat) | - | SaaS (free & Pro) | Major updates to Le Chat interface and features.[24] |
| 2025-02-06 | Le Chat (mobile) | iOS & Android apps | - | Apps + enterprise private deploys (announced) | Mobile availability and Pro tier noted.[25] |
| 2025-05-07 | Mistral Medium 3 | Text/vision LLM (enterprise) | n/a | API / partner clouds | Cost-efficient SOTA-class mid-size model.[26][27] |
| 2025-06-10 | Magistral Small | Reasoning LLM (open) | ~24B | Apache-2.0 (open weights) | First reasoning model family; open variant.[28][29] |
| 2025-06-10 | Magistral Medium | Reasoning LLM (proprietary) | n/a | API / Le Chat | Enterprise reasoning model; higher capability tier.[28][30] |
| 2025-07–09 | Voxtral (Small/Mini) | Audio-input / transcription models | - | Open weights / API | Audio input and transcription models in 2025 docs.[16] |
| 2025-07/08 | Codestral 2508 | Code LLM (FIM, test gen) | - | API | Updated code model with 256k context.[16] |
La Plateforme is Mistral AI's API service, which allows developers to integrate its commercial models into their own applications. The platform offers different pricing tiers based on model choice and usage, providing a scalable solution for businesses. Features include pay-per-token pricing model, fine-tuning capabilities, multiple deployment options (cloud, on-premises, edge), and support for both instruction-following and fill-in-the-middle code generation.[31]
Mistral's assistant product Le Chat (pronounced /lə ʃa/ in French, meaning "the cat") is delivered as a web app and (since February 2025) on iOS and Android. Launched in February 2024, it provides users with direct access to Mistral's models. The company positions Le Chat for search, coding, vision, and enterprise workflows, with a Pro tier for advanced models and features. Key features include:
Mistral distributes models through its own platform and via cloud partners, including Microsoft Azure (Azure AI model catalog), with Mistral Large first available on Azure in February 2024; the partnership attracted EU competition scrutiny.[10][11][12] Mistral Medium 3 is also published through partner marketplaces such as GitHub Models and AWS's Amazon SageMaker.[17][18]
Mistral AI's strategy is built on a dual approach:
The company has emphasized computational efficiency in its model architecture, often achieving performance comparable to larger, more resource-intensive models from competitors. This focus on efficiency makes their models more accessible and cost-effective to deploy. The company's mission is to "put frontier AI in the hands of everyone" and democratize access to AI technology, with particular emphasis on European digital sovereignty, privacy-first approach with on-premises deployment options, and multilingual support (particularly European languages).[1]
Mistral's open-weight releases emphasize efficient inference and permissive licensing (Apache-2.0). Notable innovations include:
Mistral models utilize sliding window attention (SWA) mechanism where each layer attends to the previous 4,096 hidden states, enabling efficient processing of longer sequences while reducing memory requirements.[6]
Mixtral models employ sparse mixture of experts architecture where a router network selects 2 out of 8 experts to process each token, enabling the model to have 56B total parameters while only using 13B active parameters per forward pass. This architecture balances quality and cost.[8]
Codestral Mamba uses the Mamba2 architecture, a state space model offering linear time inference and the ability to model sequences of infinite length, particularly beneficial for code generation tasks.[22]
Later offerings (Magistral) focus on reasoning-centric behavior with majority-vote decoding options.[28]
| Date | Round | Amount | Lead Investor | Valuation |
|---|---|---|---|---|
| June 2023 | Seed | €105 million ($117 million) | Lightspeed Venture Partners | Unknown |
| December 2023 | Series A | €385 million ($415 million) | Andreessen Horowitz | ~$2 billion |
| February 2024 | Series A Extension | €15 million ($16.3 million) | Microsoft | ~$2 billion |
| June 2024 | Series B | €600 million ($640 million) | General Catalyst | €5.8 billion ($6.2 billion) |
| September 2025 | Series C | €1.7 billion ($2 billion) | ASML | €11.7 billion ($13.8 billion) |
Notable investors include Lightspeed Venture Partners, Andreessen Horowitz, Nvidia, Samsung Venture Investment Corporation, Salesforce Ventures, DST Global, Index Ventures, Bpifrance, IBM, Eric Schmidt, and French tech entrepreneur Xavier Niel. ASML became the largest shareholder (~11%) after the Series C round.[4][5][9][13][14][15][2]
Press reports and company statements cite enterprise traction across European corporates and global partners:
Mistral positions its open-weight options for on-premises deployment and data-sovereign settings, while its proprietary models target higher accuracy and reasoning performance via hosted APIs. Enterprise customers include BNP Paribas, Orange, and others.[3][16]
Mistral's February 2024 investment and cloud partnership with Microsoft drew attention from the European Commission as part of broader reviews of Big Tech's AI alliances. Both companies stated they would cooperate with information requests. The UK's Competition and Markets Authority (CMA) cleared the partnership in May 2024, finding Microsoft did not have "material influence" over Mistral.[11][40][41]
Mistral AI has faced several controversies:
Despite these issues, Mistral AI continues to emphasize its European roots and open approach to AI development.