COMPANY PROFILE / AI & MACHINE LEARNING
Mistral AI
A Paris-based model developer with both open-weight and commercial models, now building compute infrastructure inside Europe
Mistral AI matters in AI and machine learning because it develops near-frontier large models in Europe, publishes their weights (some under conditions), and is also trying to own its compute base. According to its 8 September 2026 announcement, it raised EUR 3 billion in a Series D at a post-money valuation of more than EUR 21 billion, led by Samsung Electronics. Against the September 2025 Series C (EUR 1.7 billion at EUR 11.7 billion, led by ASML), the valuation in official announcements is about 1.8 times higher a year later (our calculation). On 6 October it opened a preview API of Mistral Large 4 (1 trillion total and 52 billion active parameters, with image input) and said weights would follow at the end of October (not yet confirmed). By contrast, revenue, profit and headcount could not be confirmed from official material and are not disclosed. Expanding compute (1 gigawatt by 2030) and the collaborations with Microsoft and HUMAIN are all plans or agreements in company announcements, not results. This profile sticks to the company's own official material available by 15 October 2026 (official news, model cards, documentation and the European Commission page).
- Mistral AI in 30 seconds, and where it stands in AI and machine learning
- The logo and how the name is written
- Where the company sits in AI and machine learning
- Scale
- Resources committed to AI and machine learning
- What it has done so far
- What comes next
- What the company is aiming for
- Partnership map
- Technologies, products and services
- The risks this company carries
- Glossary, references and claim-to-source audit
1. Mistral AI in 30 seconds, and where it stands in AI and machine learning
(15 rue des Halles)Opened a Munich hub in September 2026 (official news)
Cloud & infrastructure (secondary)Offering and expanding its European compute platform "Mistral Compute"
valuation over EUR 21 billionCompany announcement of 8 September 2026; led by Samsung Electronics (post-money)
Medium 3.5, Small 4Large 4 previewed on 6 October 2026; weights due at the end of October
and othersVaries by model. Medium 3.5 excludes companies with monthly revenue above US$20 million
Mistral AI promotes "sovereign AI". In the company's account this means data stays within the organization, models can be adapted, compute is private and predictable, and systems in production can be audited. Its means is owning three things: open-weight models, compute infrastructure and products (Series D announcement). For example, it released Mistral Large 3 (675 billion total and 41 billion active parameters) under Apache 2.0 in December 2025, and says Large 4 in October 2026 was trained on 3,800 NVIDIA GPUs in its own data centers in Europe Sourced. But not everything is released on the same terms. Medium 3.5 is under a "Modified MIT" license under which companies with global consolidated monthly revenue above US$20 million cannot exercise the rights (a commercial license is by separate request). The Large 4 license will be settled when weights are released. That "open-weight" carries different conditions per model is the key point in reading this company.
2. The logo and how the name is written
A note on naming: the company name is Mistral AI (the logo on the official site shows the word "Mistral" and a stepped "M" mark). The legal notice names the publisher as "Mistral, a simplified joint stock company", with share capital of EUR 15,000 and Paris register number 952 418 325 (from the website's legal notice; I did not check whether that capital is the current figure). Product names are the chat assistant Le Chat, the agentic Vibe (unified in May 2026), the developer platform Studio and the compute platform Mistral Compute. Models are named by family plus generation, such as "Mistral Large 4", "Mistral Small 4" and "Ministral 3". The logo image is the logo (SVG) used in the header of the official website mistral.ai.
Official site (brand): https://mistral.ai/brand/3. Where the company sits in AI and machine learning
As Fig. 1 shows, Mistral AI is mainly a model developer. Beyond large language models (Large, Medium, Small, Ministral) it has specialised models for document reading (OCR), speech (Voxtral), code (Devstral, Leanstral), safety classification (Shieldstral) and physics simulation (Physics AI). The other pillar is compute infrastructure: it announced Mistral Compute in June 2025 and in August 2026 made generally available "Regional Endpoints", which let customers choose Europe or the U.S. for inference. The company writes that most customers run its models inside their own data centers and cloud environments, so its "cloud" does not sit in the same position as the hyperscalers' and instead sells European sovereignty and regulatory fit Sourced. See also our explainers on Transformers and large language models, model efficiency, the EU AI Act, and AI data center power and cooling.
4. Scale
Mistral AI is unlisted. It publishes funding and valuation, but revenue, profit and headcount could not be confirmed. The confirmed figures are shown below Sourced.
| Item | What could be confirmed | Source |
|---|---|---|
| Series D (8 September 2026) | EUR 3 billion at a post-money valuation of more than EUR 21 billion. Led by Samsung Electronics, with Scaleup Europe Fund (managed by EQT) and existing investor PSG Equity as co-leads. New investors include Advent, funds managed by BlackRock and the Grand Duchy of Luxembourg; existing investors a16z, ASML, Bpifrance, NVIDIA, Salesforce Ventures and others also took part | Reference 1 |
| Series C (9 September 2025) | EUR 1.7 billion at a post-money valuation of EUR 11.7 billion. Led by ASML, with existing investors DST Global, Andreessen Horowitz, Bpifrance, General Catalyst, Index Ventures, Lightspeed and NVIDIA | Reference 2 |
| Change in valuation | About 1.8 times (21.0 / 11.7) from EUR 11.7 billion to over EUR 21 billion; both are company-announced post-money valuations and the second is a lower bound ("more than") | Our calculation |
| Business reach | "Operates across 20 countries and supports 125+ global enterprises", including Airbus, ASML and HSBC (company claim) | Reference 1 |
| Headcount | Not disclosed. 100th employee on 30 September 2024 (company page). More than 30 researchers and engineers from Emmi AI to join in May 2026 (official news) | Reference 13, 17 |
| Revenue and profit | Not disclosed (not stated in the official material I checked; press figures are not used) | Not confirmed |
A valuation is the price investors put on shares, not evidence of revenue or profit. The company calls the round "the largest equity fundraising round ever completed by a European technology company", but that compares round sizes and is not a measure of business scale. The official announcements also give no breakdown of how much of the money goes to compute, research or staff Not yet confirmed.
5. Resources committed to AI and machine learning
For each of the four resources, the table shows what could and could not be confirmed.
| Resource | What primary information confirms | Type |
|---|---|---|
| People | Current headcount is not disclosed. 100th employee on 30 September 2024; operating in 20 countries at the Series D in 2026. More than 30 researchers and engineers from Emmi AI to join in May 2026. A Munich hub with Physics AI and industrial AI research teams opened on 28 September 2026 | Sourced / current headcount not disclosed |
| Money | Funding rounds are listed on the company page only by date for the seed (13 June 2023), Series A (11 December 2023) and Series B (11 June 2024), without amounts. Series C was EUR 1.7 billion and Series D EUR 3 billion. R&D spend and capex are not disclosed. Figures such as bank loans that appear in the press are not used because they are not in the official material I checked | Sourced / spending not disclosed |
| Investments, acquisitions, JVs, divestitures | In May 2026 it announced a definitive agreement to acquire Emmi AI, an Austrian physics-AI company (completion not confirmed). On the receiving side, NVIDIA and Salesforce Ventures are among the investors. No divestitures are stated | Sourced / completion not confirmed |
| Universities, research institutes, public funds | As public investors, Bpifrance (the French public investment bank) and the Grand Duchy of Luxembourg are in the Series D. For the environmental study it worked with the French ecological transition agency ADEME and Carbone 4, with peer review by Resilio and Hubblo (July 2025). University collaborations could not be confirmed | Sourced / universities not confirmed |
Next, the status of the main models and compute facilities. Stages and figures follow the company's own wording, and performance is the developer's stated value.
| Model / facility | Scale | Stage | Date | Source |
|---|---|---|---|---|
| Mistral Large 4 | 1 trillion total / 52 billion active parameters, natively multimodal. Trained on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers (company description) | Preview (API) / weights announced Not yet confirmed | 6 October 2026 (weights announced for end of October) | Ref. 3 |
| Mistral Large 3 | 675 billion total / 41 billion active parameters (MoE). Trained on 3,000 NVIDIA H200 GPUs. Base and instruct versions released under Apache 2.0 | Released (weights) | 2 December 2025 | Ref. 4 |
| Mistral Medium 3.5 | Dense 128 billion parameters, 256k-token context, image input. Modified MIT license. The company says it can be self-hosted on as few as four GPUs | Released (weights) | 22 May 2026 | Ref. 6, 7 |
| Mistral Small 4 | 119 billion total / 6 billion active parameters (MoE, 4 of 128 experts active), 256k-token context. Apache 2.0 | Released (weights) | 16 March 2026 | Ref. 5 |
| Mistral Compute | Planned to offer "tens of thousands of GPUs" on NVIDIA reference architectures and expand (June 2025 announcement). In July 2026 Microsoft announced a multibillion-dollar commitment (value undisclosed) to use part of Mistral's European GPU capacity, with Mistral adding thousands of NVIDIA Vera Rubin GPUs | Partly operating, expanding Not yet confirmed | June 2025, July 2026 | Ref. 9, 11 |
| European compute capacity | "Mistral will build one gigawatt of European compute capacity by 2030" (official news, 28 September 2026). In August 2026 it announced "European Compute Units (ECUs)" to pool multi-year enterprise commitments | Target Not yet confirmed | August, September 2026 | Ref. 8, 15 |
The GPU counts are for different generations and different purposes (3,000 H200, 3,800 Grace Blackwell, "thousands" of Vera Rubin). They are numbers used for training, not the total the company owns, and are neither added up nor converted into capacity (MW). Mistral's total GPU count and total power are not stated in the official material I checked Not yet confirmed.
6. What it has done so far
| When | What happened | Meaning |
|---|---|---|
| April-December 2023 | Founded (April), first employee (5 June), seed (13 June), Mistral 7B (27 September), Series A (11 December) | First public model about half a year after founding |
| February-September 2024 | Mistral Large (26 February), Series B (11 June), 100th employee (30 September) | Added a commercial top-tier model |
| January-June 2025 | Small 3 (30 January), the new Le Chat (6 February), Mistral OCR (6 March), Mistral Code (4 June), Mistral Compute (11 June) | Widened from models to compute |
| 22 July 2025 | Published a lifecycle assessment of an AI model (environmental impact) | The company published figures on the environmental load of training and inference (chapters 5 and 10) |
| 9 September 2025 | Series C (EUR 1.7 billion, EUR 11.7 billion valuation, led by ASML) | Strategic partnership with chip-equipment maker ASML |
| October-December 2025 | Mistral Studio (24 October), partnerships with SAP and Helsing in Germany (19 November), Mistral 3 (2 December; Large 3 under Apache 2.0), Devstral 2 and Vibe CLI (9 December) | Released a large MoE's weights under a permissive license |
| March 2026 | Founding member of the NVIDIA Nemotron Coalition (16 March), Small 4 (16 March), Forge (17 March; a system for enterprises to build models on their own knowledge) | A plan to co-develop open frontier models with NVIDIA |
| May 2026 | Medium 3.5 and Vibe remote agents (22nd), the Emmi AI agreement (23rd), the Physics AI launch (27th) | Focus on coding and industrial use |
| 21 July 2026 | Expanded partnership with Microsoft: a multibillion-dollar commitment by Microsoft to use Mistral's European GPU capacity, with Medium 3.5 and OCR 4 available in Foundry and Copilot Studio | The character of a compute seller |
| August 2026 | Regional Endpoints generally available and Priority Tier in preview, the ECU concept (11th), collaboration with HUMAIN (24th) | Regional choice in Europe and the U.S., and long-term capacity |
| 8 September 2026 | Series D (EUR 3 billion, valuation over EUR 21 billion, led by Samsung Electronics) | Described by the company as the largest equity round by a European technology company |
| 6 October 2026 | Mistral Large 4 preview (weights announced for end of October) | A trillion-parameter-class open-weight announcement |
7. What comes next
October 2026Until then, red-teaming with vetted partners. License not yet announced
by 2030Company target (official news, 28 September 2026). No interim stages or sites stated
Vera Rubin GPUsPlanned addition. Value, count and timing undisclosed (joint announcement)
cybersecurity and voiceDescribed as "hundreds of millions of Euros"; the notice says it depends on subsequent commercial agreements
The above are plans, targets and agreements in company announcements, not results Not yet confirmed. The HUMAIN announcement states that it contains forward-looking statements and that actual results may differ because of uncertainties including subsequent commercial agreements.
Revenue, profit and headcount / R&D spend and capex / the total number of GPUs owned and total power / data center locations and capacity already operating (where official material is silent) / completion of the Emmi AI acquisition / the Large 4 license and the actual release of weights / value, count and timing of the Microsoft agreement / funding figures such as bank loans that appear in the press (not adopted because they are not in official material).
8. What the company is aiming for in AI and machine learning
This separates what the company says directly from what we read into it. Inference marks our own reading of primary information.
| What can be read | Primary information behind it | Type |
|---|---|---|
| It makes "sovereign AI" the central idea of its business | The Series D announcement defines it as controlling data, models, compute and production systems, and claims to be "the only AI company in the world building the full stack" | Sourced (company claim) |
| It targets European customers and governments first | SAP and Helsing in Germany, the Munich hub, the European Compute Units (with Amadeus, ASML, Capgemini, Caisse des Dépôts and CMA CGM quoted), and European capacity with Microsoft | Sourced / Inference |
| It is widening into industry and engineering (physics AI) | The Emmi AI agreement, the Physics AI launch, the Munich research teams, and statements naming semiconductors, automotive and aerospace | Sourced / Inference |
| It builds frontier models in close alignment with NVIDIA | Founding member of the NVIDIA Nemotron Coalition, NVIDIA GPUs for training Mistral 3 and Large 4, NVIDIA as a Series C and D investor, and released NVFP4 weights | Sourced / Inference |
| It balances openness and commercial use through per-model conditions | Apache 2.0 (Large 3, Small 4, Ministral 3), a Modified MIT excluding companies above a revenue threshold (Medium 3.5), and "Open" shown for Large 4 | Sourced / Inference |
The basis for reading it as balancing openness and commercial use per model is that licenses have changed with generations. Mistral 3 (December 2025), including Large 3, was Apache 2.0, but Medium 3.5 (2026) moved to a Modified MIT under which companies with global consolidated monthly revenue above US$20 million cannot use it (license text). Large 4 is shown only as "Open" in the documentation, and the terms will not be known until release. It looks like a design that opens smaller models widely while steering high-revenue companies to commercial licenses or hosting. However, the company does not state that intent, and the reason for the change could not be confirmed Not yet confirmed
9. Partnership map in AI and machine learning
As the figure shows, NVIDIA plays three roles at once: investor, GPU supplier and co-development partner. ASML and Samsung Electronics are major makers of chip-making equipment and chips respectively, and the company announced both investment and strategic partnership. Microsoft is a distribution channel (Foundry, Copilot Studio) and also a user of Mistral's European capacity (value undisclosed). The company describes the HUMAIN collaboration as "hundreds of millions of Euros", with the forward-looking notice described above Sourced.
10. Technologies, products and services
Model performance and figures are the developer's (company's) stated values. Benchmarks are given where name and setting can be confirmed, and no ranking against other companies' models is made. The company's comparisons use rival models and indices it chose.
| Item | Official statement (developer's stated value) | What could not be confirmed |
|---|---|---|
| Model families and licenses | Large 3, Small 4 and Ministral 3 (3B / 8B / 14B): Apache 2.0. Medium 3.5: Modified MIT (excluding companies above US$20 million monthly revenue). Leanstral 1.5 and Shieldstral: Apache 2.0 (Hugging Face listings). Voxtral TTS weights: CC BY-NC 4.0 (same). Large 4: "Open" (documentation) | The Large 4 license text; the Devstral 2 license text (listed as "other" on Hugging Face) |
| Benchmarks (company-reported) | Medium 3.5: 77.6% on SWE-Bench Verified and 91.4% on τ³-Telecom (model card; settings not checked here). Large 4: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4.0, 93% on Cybench, and a top-five rank on the Artificial Analysis Cyber Index (company text; includes a claim of "state-of-the-art among open models") | Independent reproduction; differences between the Large 4 preview and the final release |
| Products | Le Chat (chat), Vibe (agent with Work and Code modes, VS Code extension, remote agents), Studio (managing prompts, skills and connectors), Forge (enterprises building models on their own knowledge), Workflows, Search Toolkit and Agentic Search. Plans: Free, Pro (US$14.99 a month), Team (US$24.99 per user a month) and Enterprise (pricing page, 15 October 2026) | User numbers and revenue |
| Compute | Mistral Compute (from bare-metal servers to managed PaaS), Regional Endpoints (inference in Europe or the U.S., generally available), Priority Tier (with an uptime SLA, in preview), and hosting of third-party open models (Z.ai GLM-5.2 and 5.3) | Facility locations and capacity; uptime record |
| Environmental impact (company study) | Mistral Large 2: as of January 2025, training plus 18 months of use gave 20.4 ktCO2e, 281,000 m3 of water and 660 kg Sb eq of resource depletion. One 400-token Le Chat response gave 1.14 gCO2e, 45 mL of water and 0.16 mg Sb eq (excluding users' terminals). Done with Carbone 4 and ADEME and peer-reviewed by Resilio and Hubblo; the company calls it "a first approximation" | Equivalent figures for models since 2025 (Large 3 and 4) |
Mistral's official material repeatedly describes ways to get the same performance from smaller equipment. Large 3 has 675 billion total parameters but keeps active parameters to 41 billion through MoE, and released NVFP4 (4-bit floating point) weights so it can run on "Blackwell NVL72 systems or a single 8xA100 or 8xH100 node". Small 4 has 119 billion total and 6 billion active, and Medium 3.5 is a dense 128 billion that the company says can be self-hosted on as few as four GPUs (all company claims). On the training side, Mistral 3 used NVIDIA Hopper GPUs with HBM3e (high bandwidth memory) and Large 4 moved to Grace Blackwell, so each GPU generation changes the training equipment and power design directly Sourced. On power, the company's own 2025 study counts upstream impacts from manufacturing servers as well and concludes that "impacts are roughly proportional to model size". The 1-gigawatt-by-2030 target is a figure that runs into the power equipment, cooling and transmission constraints of this field Inference. See our explainers on model efficiency, inference optimization and serving, AI accelerators, HBM, and AI data center power and cooling.
11. The risks this company carries
The following is arranged neutrally, within what the company's official material and the European Commission page say.
| Risk | Details | Type |
|---|---|---|
| Capital intensity and opaque earnings | Expanding compute (1 GW by 2030) and training frontier models take large sums. The valuation exceeds EUR 21 billion, but revenue, profit and spending are undisclosed, so the prospects of recovering the investment cannot be checked from outside | Not yet confirmed |
| Dependence on NVIDIA and overlapping interests | Training uses NVIDIA H200 and Grace Blackwell, with Vera Rubin planned for additions. NVIDIA is also an investor and a co-development partner (company announcements). Supply constraints or price changes feed directly into plans | Sourced / Inference |
| Differences in license terms | Terms differ by model: Apache 2.0, Modified MIT (excluding companies above US$20 million monthly revenue), CC BY-NC 4.0, and Large 4 undecided. Users must check each model's text | Sourced |
| Releasing a highly capable model | The company describes Large 4 as strong in cybersecurity and says that before weights are released, vetted partners and state authorities will red-team a version with reduced moderation and expanded cyber capabilities. Details of assessments of potential misuse after release are not in the company's material (this article gives no specific procedures) | Sourced (company statement) |
| Self-reported evaluations | The Large 4 comparisons use models and indices the company chose and include phrases such as "state-of-the-art among open models" and "significantly outperforming" certain open-weight models. Third-party evaluations (Artificial Analysis, Surge AI human ratings) are cited by the company, and independent reproduction could not be confirmed | Sourced (company claim) |
| EU AI regulation | According to the European Commission page, Mistral AI is listed as a signatory of the General-Purpose AI Code of Practice, a voluntary tool (checked 15 October 2026). The code is a way to show compliance with AI Act obligations, and signing is not itself a guarantee of compliance. How the AI Act applies depends on a model's capability class and how it is released (Regulation (EU) 2024/1689) | Sourced |
| Contingency of partnerships | The Microsoft agreement has no disclosed value or timing, the HUMAIN collaboration carries a notice that it depends on subsequent commercial agreements, and the ECU concept is at the stage of gathering multiyear commitments (official news) | Not yet confirmed |
| Limits of disclosure | Headcount, revenue, profit, owned GPUs and facility locations are undisclosed. Whether the share capital shown in the legal notice (EUR 15,000) is current could not be confirmed | Not yet confirmed |
This article summarises a company and is not legal, safety or procurement advice. Use of the models is subject to each license and to national and regional law.
12. Glossary
- Sovereign AI
- Being able to control data, models, compute and production systems yourself, as an organization or country (Mistral AI's definition).
- Open-weight
- Publishing the trained model's weights. The conditions of use depend on the license.
- Mixture of experts (MoE)
- A design that activates only some "experts" per input; total and active parameters differ.
- Dense model
- A design in which all parameters work every time. All 128 billion are active in Medium 3.5.
- NVFP4
- NVIDIA's 4-bit floating-point format; makes weights smaller and cuts needed memory.
- ECU
- European Compute Units: a scheme converting enterprises' multiyear commitments into access to compute infrastructure Mistral builds.
- Physics AI
- Models that predict the behavior of physical systems; the company describes using them to replace or assist simulation.
- GPAI Code of Practice
- A voluntary tool to help meet the EU AI Act's obligations for general-purpose AI models.
- LCA (lifecycle assessment)
- An assessment of environmental impact from manufacturing through use.
13. References
- Mistral AI Official news, "Mistral raises €3B to make sovereign, open-weight AI the technology frontier" (8 September 2026). Series D, valuation, investors, business reach. https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/
- Mistral AI Official news, "Mistral raises 1.7B€ to accelerate technological progress with AI" (9 September 2025). Series C. https://mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai/
- Mistral AI Official news, "Introducing Mistral Large 4" (6 October 2026). 1 trillion parameters, 3,800 Grace Blackwell GPUs, benchmarks, announced weight release. https://mistral.ai/news/mistral-large-4/
- Mistral AI Official news, "Introducing Mistral 3" (2 December 2025). Large 3, Ministral 3, Apache 2.0, 3,000 H200 GPUs. https://mistral.ai/news/mistral-3/
- Mistral AI Official news, "Introducing Mistral Small 4" (16 March 2026). https://mistral.ai/news/mistral-small-4/
- Mistral AI Official news, "Remote agents in Vibe. Powered by Mistral Medium 3.5." (22 May 2026). https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/
- Mistral AI Hugging Face model card and license, "Mistral-Medium-3.5-128B". Modified MIT (excluding companies above US$20 million monthly revenue), benchmarks. https://huggingface.co/mistralai/Mistral-Medium-3.5-128B
- Mistral AI Official news, "In-region inference, open models, and new European infrastructure for sovereign AI." (11 August 2026). Regional Endpoints, Priority Tier, ECUs. https://mistral.ai/news/regional-inference-open-models-new-compute/
- Mistral AI Official news, "Mistral Compute" (11 June 2025). https://mistral.ai/news/mistral-compute/
- Mistral AI Official news, "Mistral partners with NVIDIA to accelerate open frontier models" (16 March 2026). https://mistral.ai/news/mistral-ai-and-nvidia-partner-to-accelerate-open-frontier-models/
- Microsoft / Mistral Joint announcement, "Microsoft and Mistral expand strategic partnership..." (PR Newswire, 21 July 2026). https://tools.prnewswire.com/en-us/live/20813/release/20260721EN08769
- Mistral AI Official news, "Mistral x HUMAIN" (24 August 2026). https://mistral.ai/news/mistral-x-humain/
- Mistral AI Official news, "Emmi joins Mistral to accelerate the AI-native industry" (23 May 2026). https://mistral.ai/news/accelerate-ai-native-industry/
- Mistral AI Official news, "Mistral - KI für Deutschland" (19 November 2025). SAP, Helsing. https://mistral.ai/news/ki-fur-deutschland/
- Mistral AI Official news, "Hallo, Deutschland!" (28 September 2026). The Munich hub and 1 GW by 2030. https://mistral.ai/news/hallo-deutschland/
- Mistral AI Official news, "Our contribution to a global environmental standard for AI" (22 July 2025). Environmental impact figures. https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai/
- Mistral AI Company page (checked 15 October 2026). Founding, key dates, founders. https://mistral.ai/company
- Mistral AI Legal notice. Publisher, share capital, register number, address. https://mistral.ai/legal/
- Mistral AI Official documentation, "Models Overview" (checked 15 October 2026). Model list and licenses. https://docs.mistral.ai/getting-started/models/models_overview/
- European Commission "The General-Purpose AI Code of Practice" (list of signatories, checked 15 October 2026). https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
- Mistral AI Official news index (checked 15 October 2026). https://mistral.ai/news
- Mistral AI Official Hugging Face organization "mistralai" (retrieved 15 October 2026). Model and license listings. https://huggingface.co/mistralai
- Mistral AI Pricing page (checked 15 October 2026). https://mistral.ai/pricing
- Mistral AI Official brand page (logo source; the logo is the SVG in the site header). https://mistral.ai/brand/
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Name, legal notice (simplified joint stock company, register number, address), founding (April 2023), founders, key dates, 100th employee | Sourced | Reference 17 https://mistral.ai/company Reference 18 https://mistral.ai/legal/ |
| Series D (EUR 3 billion, valuation over EUR 21 billion, investors, 20 countries and 125+ enterprises) | Sourced | Reference 1 https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/ (includes company claims) |
| Series C (EUR 1.7 billion, EUR 11.7 billion valuation, led by ASML, investors) | Sourced | Reference 2 https://mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai/ |
| Mistral Large 4 (1T / 52B, 3,800 Grace Blackwell GPUs, benchmarks, announced weight release, red-teaming) | Sourced | Reference 3 https://mistral.ai/news/mistral-large-4/ (includes company claims) |
| Mistral 3 (Large 3 at 675B / 41B, 3,000 H200 GPUs, Apache 2.0, NVFP4) | Sourced | Reference 4 https://mistral.ai/news/mistral-3/ |
| Small 4 (119B / 6B, Apache 2.0) | Sourced | Reference 5 https://mistral.ai/news/mistral-small-4/ |
| Medium 3.5 (dense 128B, 256k tokens, Modified MIT, excluded above US$20 million monthly revenue, four GPUs, SWE-Bench Verified) | Sourced | Reference 6 https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/ Reference 7 https://huggingface.co/mistralai/Mistral-Medium-3.5-128B (company's stated values) |
| Regional Endpoints, Priority Tier, ECUs, hosting of third-party open models | Sourced | Reference 8 https://mistral.ai/news/regional-inference-open-models-new-compute/ |
| Mistral Compute (June 2025, NVIDIA reference architectures, plan for tens of thousands of GPUs, launch partners) | Sourced | Reference 9 https://mistral.ai/news/mistral-compute/ |
| Founding member of the NVIDIA Nemotron Coalition | Sourced | Reference 10 https://mistral.ai/news/mistral-ai-and-nvidia-partner-to-accelerate-open-frontier-models/ |
| The expanded Microsoft partnership (multibillion-dollar commitment, thousands of Vera Rubin GPUs, Foundry and Copilot Studio) | Sourced | Reference 11 https://tools.prnewswire.com/en-us/live/20813/release/20260721EN08769 (value, count and timing undisclosed) |
| Collaboration with HUMAIN (hundreds of millions of euros, forward-looking notice) | Sourced | Reference 12 https://mistral.ai/news/mistral-x-humain/ |
| The Emmi AI agreement (30+ people) | Sourced | Reference 13 https://mistral.ai/news/accelerate-ai-native-industry/ |
| Partnerships with SAP and Helsing, the Munich hub, the 1 GW by 2030 target | Sourced | Reference 14 https://mistral.ai/news/ki-fur-deutschland/ Reference 15 https://mistral.ai/news/hallo-deutschland/ |
| Environmental study (20.4 ktCO2e, 281,000 m3, 1.14 g, 45 mL, ADEME and Carbone 4) | Sourced | Reference 16 https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai/ |
| Model list and license labels (Large 4 as "Open"), Z.ai GLM models | Sourced | Reference 19 https://docs.mistral.ai/getting-started/models/models_overview/ Reference 22 https://huggingface.co/mistralai |
| Listing as a signatory of the GPAI Code of Practice | Sourced | Reference 20 https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai |
| Pricing (Free, Pro, Team) | Sourced | Reference 23 https://mistral.ai/pricing |
| Logo image (SVG in the official site header) | Sourced | Reference 24 https://mistral.ai/brand/ |
| Change in valuation (EUR 11.7 billion to over EUR 21 billion, about 1.8 times) | Our calculation | our own: 21.0 / 11.7 (lower bound) |
| Revenue, profit, headcount, R&D spend, capex, total GPUs and power, completion of Emmi AI, the Large 4 license, the value of the Microsoft agreement | Not yet confirmed | Not stated in the official material checked. Not estimated in this article (items not confirmed) |
| Release of Large 4 weights (end of October 2026), 1 GW by 2030, thousands of Vera Rubin GPUs, ECUs and HUMAIN plans | Not yet confirmed | Reference 3 https://mistral.ai/news/mistral-large-4/ Reference 8 https://mistral.ai/news/regional-inference-open-models-new-compute/ Reference 11 https://tools.prnewswire.com/en-us/live/20813/release/20260721EN08769 Reference 12 https://mistral.ai/news/mistral-x-humain/ Reference 15 https://mistral.ai/news/hallo-deutschland/ Company targets and plans, not results |
| The readings on sovereign AI, European customers, industrial AI, closeness to NVIDIA and license balancing | Inference | Reference 1 https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/ Reference 3 https://mistral.ai/news/mistral-large-4/ Reference 7 https://huggingface.co/mistralai/Mistral-Medium-3.5-128B Reference 8 https://mistral.ai/news/regional-inference-open-models-new-compute/ Reference 13 https://mistral.ai/news/accelerate-ai-native-industry/ our own interpretation |
| The explanation linking model size, precision and GPU generation to equipment and power, and the explanation of constraints on 1 GW | Not yet confirmed | our own explanation (general commentary, not a company disclosure) |
Last updated: 15 October 2026 / Troy Technical
The figures in this article come from Mistral AI's official news, model cards, documentation and legal notice, the joint announcement with Microsoft, and the European Commission page. Model names, prices, contracts and corporate structure were checked on 15 October 2026 and may change. Market-research estimates and press-report-only figures are not used. This is not an assessment of the company's financial condition or of the relative merit of its products. Passages marked "Inference" are our own reading of primary information and are not statements by the company.