When people talk about the big AI labs, it is usually the same names from the US and China. Mistral AI, based in Paris, is the most serious European answer, and the company is growing at a pace Europe has not seen in a long time.
I've followed Mistral since the first Mixtral release, because the lab combines two things that rarely go together: strong open models and a consistently European setup, from GDPR compliance to its own data center south of Paris.
I've pulled together the current numbers on Mistral here: valuation, funding, revenue, Le Chat usage, the full model family, benchmarks and, of course, where its data centers are located.
- Mistral AI was founded in 2023 in Paris and valued at €11.7B in its Series C (Sep. 2025). In June 2026, talks of around €20B were reported
- Annualized revenue jumped from around $16M (end of 2024) to over $400M (Jan. 2026), with a target of €1B by the end of 2026
- Mistral's first dedicated data center is in Bruyères-le-Châtel (Essonne) near Paris, with 13,800 Nvidia GB300 GPUs and 44 MW, financed through $830M of debt
- With Le Chat (around 5M users) and a broad open-weights family, Mistral is Europe's most important AI alternative to OpenAI, Anthropic and Google
1. Mistral AI at a glance
Mistral AI was founded in spring 2023 in Paris by Arthur Mensch (previously Google DeepMind), Guillaume Lample and Timothée Lacroix (both previously Meta FAIR). In just over two years, it has become Europe's most valuable AI startup. The key facts:
Metric | Value |
|---|---|
| Founded | 2023, Paris |
| Founders | Arthur Mensch, Guillaume Lample, Timothée Lacroix |
| Valuation | €11.7B (Series C, Sep. 2025) |
| Annualized revenue | over $400M (Jan. 2026) |
| Le Chat users | around 5M monthly |
| Flagship model | Mistral Large 3 (675B parameters, MoE) |
2. Funding & valuation
Mistral's rise shows in its funding rounds. From the €105M seed (June 2023) it went to a €385M Series A (December 2023), a €600M Series B (June 2024) and finally a €1.7B Series C (September 2025). In total, Mistral has raised around $3B in equity:
The valuation exploded in parallel, from around €240M at the seed to €11.7B in the Series C, led by Dutch chip-equipment maker ASML. The chart below shows the jump across the rounds:
And that's not the end. In June 2026, Mistral was reported to be in talks for a new round of around €3B at a valuation of about €20B. The talks were still early at that point, so treat the number with caution.
3. Valuation in context
As impressive as Mistral's growth is from a European perspective, the gap between Mistral and the US giants is stark. OpenAI and Anthropic are valued above $850B, while Mistral sits at $13.5B (converted from €11.7B), in a completely different league:
gradually.aiThat puts the headlines in perspective. Mistral leads in Europe, but measured against the capital flowing in the US, it's still a challenger. That's exactly why the company leans on efficiency, open models and its own infrastructure rather than trying to match pure cash burn.
4. Revenue & growth
On revenue, Mistral is catching up fast. Annualized revenue (the run rate, the current monthly revenue extrapolated to a year) rose from around $16M at the end of 2024 to over $400M in January 2026, a twentyfold increase in just over a year:
CEO Arthur Mensch has set the goal of crossing €1B in revenue by the end of 2026. The main drivers are the enterprise business and Le Chat, the in-house AI assistant.
5. Le Chat & usage
Le Chat is Mistral's answer to ChatGPT and the most important consumer-facing product. After the mobile launch in early 2025, it climbed quickly:
Metric | Value |
|---|---|
| Monthly Le Chat users | ~5M |
| App downloads (first week) | 1M |
| Desktop visits mistral.ai (March 2026) | 10.8M |
Compared to ChatGPT with its hundreds of millions of users, that's still small. But for a European product without the marketing machine of a US giant, the curve is remarkable. For more numbers on the whole AI industry, see my AI statistics.
6. The model family
Mistral runs a two-track strategy: powerful Large models for the API and enterprise customers, and a broad family of open models anyone can self-host. The overview below shows the most important ones:
Model | Parameters | Context | Release | Type |
|---|---|---|---|---|
| Mistral Large 3 | 675B (MoE, 41B active) | 256K | Dec. 2025 | Open-weights |
| Mistral Large 2 | 123B | 128K | Jul. 2024 | Open-weights |
| Mistral Small 3.1 | 24B | 128K | Jan. 2025 | Open-weights |
| Mixtral 8x22B | 176B (44B active) | 65K | Apr. 2024 | Open-weights |
| Mixtral 8x7B | 56B (14B active) | 33K | Dec. 2023 | Open-weights |
| Mistral 7B | 7B | 33K | Sep. 2023 | Open-source |
| Ministral 8B / 3B | 8 / 3B | 128K | Oct. 2024 | Open-weights (edge) |
| Codestral | n/a | 32K | May 2024 | Code, proprietary |
The current flagship, Mistral Large 3 (December 2025), is a mixture-of-experts model with 675B parameters, of which only 41B are active per request. That keeps inference comparatively cheap. For a broader comparison of all major models, see my LLM statistics and open-source LLMs.
7. Benchmarks
On raw benchmarks, you have to be honest. Mistral Large 3 is a solid mid-tier open-weights model, but not a frontier model on the level of Claude Opus or GPT-5.x. On the Artificial Analysis Intelligence Index it lands mid-pack among open models:
Benchmark | Mistral Large 3 | Context |
|---|---|---|
| Artificial Analysis Intelligence Index | 16 | mid-tier among open-weights models |
| LMSYS Chatbot Arena | ~1,418 Elo | 2nd among open non-reasoning models |
| MMLU-Pro | ~85.5% | strong general knowledge |
| HumanEval | ~92% | solid coding |
| GPQA Diamond | ~43.9% | below the reasoning frontier |
The figures come from third parties like Artificial Analysis and LMSYS, since Mistral prefers other benchmarks. Concretely that means around 1,418 Elo in the LMSYS arena and about 85.5% on MMLU-Pro, so strong knowledge, plus a solid 92% on HumanEval. On hard reasoning (GPQA Diamond, only around 43.9%) Mistral sits well behind the top. For everyday and enterprise tasks that is plenty; for complex reasoning chains, a frontier model is the better pick.
8. API pricing in context
Mistral's strongest argument is price. Mistral Large 3 costs $2 per million input tokens, far less than Claude Opus and on par with Gemini. The smaller models are cheaper still:
Mistral Small 3.1 comes in at $0.20 per million tokens, one of the cheapest usable options on the market. For price-sensitive, high-volume use cases that is a real argument. For a full cost calculation across providers, see the API cost calculator.
9. Context window & knowledge cutoff
Mistral's context windows have grown sharply across generations, from 33,000 tokens in the first models to 256,000 in Mistral Large 3. The knowledge cutoff (the date up to which a model contains training data) moves forward accordingly:
At 256,000 tokens, Mistral Large 3 handles roughly 600 pages of text at once. That's enough for long documents and large code bases, but below the 1 million tokens offered by Gemini and the current Claude and GPT models.
10. Data centers & infrastructure
Mistral's infrastructure strategy is the part worth watching. For a long time, it relied on rented cloud capacity like almost every AI lab. That's changing. Mistral is building its first dedicated data center, deliberately in France:
Feature | Value |
|---|---|
| Location | Bruyères-le-Châtel (Essonne, ~30 km south of Paris) |
| Operator | Eclairion |
| Hardware | 13,800 Nvidia GB300 (Grace Blackwell) |
| Power | 44 MW |
| Go-live | Q2 2026 |
| Financing | $830M (debt, March 2026) |
| Europe target | 200 MW by end of 2027 |
The site is Bruyères-le-Châtel in the Essonne department, around 30 km south of Paris, operated by the French firm Eclairion on a roughly 4-hectare plot at the edge of the castle grounds. It is stocked with 13,800 Nvidia GB300 GPUs and delivers 44 MW. Mistral financed the build through an $830M debt round, the largest AI debt raise by a European tech company to date.
And that's only the start. Mistral has announced plans to build out around 200 MW of compute across several European sites by the end of 2027:
Why it matters: owning infrastructure on European soil makes Mistral less dependent on US cloud providers and is a strong selling point for European governments and enterprises that do not want their data sitting in the US.
11. Open source & European AI sovereignty
Mistral's real differentiator is its stance. While OpenAI and Anthropic keep their best models strictly closed, Mistral has released open models from the start. Mistral 7B and the Mixtral series became standard building blocks for developers who want to self-host their AI.
Then there is the European card. Mistral aggressively markets GDPR compliance, EU AI Act readiness and now its own infrastructure in France. With ASML as a major investor and Nvidia hardware on French soil, Mistral has become the poster child for European AI sovereignty.
My take. Mistral won't overtake OpenAI and Anthropic on raw benchmarks any time soon. But as a cheap, open and European alternative, the lab owns a niche no US provider can serve as credibly. Anyone who values data sovereignty, open weights and fair pricing can't ignore Mistral.
