Europe's AI Landscape — Between Catching Up and Forging Its Own Path

Mistral, Aleph Alpha, Teuken, EuroLLM, Apertus: Europe has more of its own language models than most people realize. An overview of models, providers, and a comparison with the U.S. and China—and a very practical question: Which of these actually still run on your own computer?

When the board discusses AI strategy, three names usually come up first: ChatGPT, Gemini, and Claude. Perhaps DeepSeek as well, ever since the Chinese alternative caused a stir in 2025. What’s less often discussed: Europe has been building its own AI ecosystem for years—with models you should be familiar with before choosing a provider.

For companies that, like many of our clients, operate in regulated industries—pharmaceuticals, manufacturing, quality management—this is no academic question. It’s about data protection, auditability, and, increasingly, a very practical consideration: Does this query really need to be sent to a server in California or Hangzhou, or is a model running in Frankfurt, Paris, or on the company’s own server sufficient?

The Global Race: The U.S., China, Europe

Three regions, three very different strategies. The U.S. invests the largest training budgets and dominates in frontier models and consumer reach. China relies on open weights as a geopolitical tool and, with DeepSeek, Qwen (Alibaba), Kimi (Moonshot AI), and Z.ai (formerly Zhipu), cost-effective models at a world-class level. Europe is pursuing a third path: smaller but growing models, backed by a clear regulatory and legal framework.

💡 What this means in practice:

For most business applications in the SME sector—text summarization, internal search, translation, code analysis—the performance differences between a good European model and a U.S. model are rarely the deciding factor today. What often matters most is where the data is located and under which jurisdiction it is processed. 

A Comparison of the European Model Landscape

Contrary to popular belief, Europe is not a blank spot on the AI map. In addition to the well-known flagship Mistral AI, there is a whole range of models, most of which originate from publicly funded research consortia—and are freely accessible.

The commercial frontrunners:

Mistral AI (Paris, France) Frontier Lab 

Founded in 2023, it is the only European lab with consistently competitive frontier models, including Mistral Large 3 and Medium 3.5 (2026). EU hosting as standard, plus several open models (including Mistral 7B, Mixtral, and Codestral) under the Apache 2.0 license for self-hosting.

AlephAlpha / PhariaAI (Heidelberg, Germany) Sovereign Infrastructure

Has evolved from a standalone chat API to “Specialized Language Models” for the public sector, defense, and industry. Reference customers include SAP, Bosch, Deutsche Bank, and the Federal Employment Agency—the focus is on large-scale projects capable of on-premises deployment, rather than on a broad offering for end customers.     

The open, publicly funded models:

Teuken-7B / OpenGPT-X (Germany)Open Source

A German research project supported by the Fraunhofer Institutes IAIS and IIS, the Technical University of Dresden, the Jülich Research Center, and industry partners such as Aleph Alpha and IONOS, funded with approximately 15 million euros by the Federal Ministry for Economic Affairs and Energy. Trained from the ground up on all 24 official EU languages; available for free on Hugging Face since late 2024.

EuroLLM-22B (EU Consortium) Open Source

EU-funded consortium comprising the University of Lisbon, the University of Edinburgh, Université Paris-Saclay, and Unbabel. Trained from scratch for 35 languages on the MareNostrum 5 supercomputer with over 4 trillion tokens and a 32K context window; licensed under Apache 2.0—according to the development team, one of the most powerful fully open-source “Made in EU” models for multilingual applications.

Apertus (ETH Zurich / EPFL, Switzerland)Fully open-source

Trained on the Alps supercomputer in Lugano using over 10,000 Nvidia Grace Hopper superchips. Released with 70 billion parameters—including training data, intermediate checkpoints, and training recipes—under the Apache 2.0 license. According to the developers, it is one of the most transparent language models in the world.

ALIA (Barcelona Supercomputing Center, Spain) & Amália (Portugal) National

Spain produced ALIA (40 billion parameters), an open model trained at the Barcelona Supercomputing Center. Portugal followed in July 2026 with Amália, the first open LLM in European Portuguese, funded with approximately 7 million euros in public funds.

OpenEuroLLM In Development

A long-term, EU-funded initiative aimed at developing a competitive, open, multilingual European language model. Announced for 2024/2025; as of 2026, not yet ready for production.

Specialists Instead of Generalists

Not every European AI solution needs to be a general-purpose LLM. For many use cases, specialized providers deliver better results than a large language model: DeepL (DE) for translation, Black Forest Labs (DE) for image generation, LightOn (FR) for document intelligence, Kyutai (FR) for natural language processing, and H Company (FR) for agent-based automation.

How competitive is Europe really?

The honest answer: it varies, depending on the scale. In terms of pure model size and training budget, OpenAI, Google, and the Chinese providers remain ahead—partly because they have significantly greater private computing capacity. The largest multimodal models still come from the U.S., and cutting-edge research is often put into practice more quickly in California than in Paris or Heidelberg.

At the same time, the gap has narrowed significantly when it comes to business applications suitable for everyday use: Mistral’s latest models certainly compete with established U.S. models in terms of speed and quality, DeepL surpasses Google Translate’s translation quality in many language pairs, and image generation from Black Forest Labs keeps pace with international providers.

However, Europe’s real trump card lies not primarily in raw performance, but in its foundation: GDPR compliance from the outset rather than retroactive adaptation, no vulnerability to the U.S. CLOUD Act, native alignment with the EU AI Act—and, often underestimated, better support for smaller European languages, because many European models were trained to be multilingual from the ground up, rather than using English as a starting point and tacked-on other languages later.

European AI doesn’t have to be the biggest to be relevant. It must be the most trustworthy for the European legal landscape.

What other European open-source models can I run myself on an NVIDIA PC?

That’s the question that often matters most in practice—especially for companies that don’t want to send sensitive data (formulas, design plans, quality data, legacy source code) to the cloud in the first place. The good news: Thanks to quantization, several European models fit on off-the-shelf Nvidia graphics cards—ranging from models with 7 billion parameters up to 70 billion parameters, if you’re willing to work with multiple GPUs or make compromises on speed.

What hardware resources are required?

What matters is not the model size alone, but how much graphics memory (VRAM) the quantized model occupies. The standard practice is quantization to 4-bit (Q4_K_M), which reduces memory requirements by about 75% compared to the original precision—with only a minor loss in quality that is barely noticeable for most business applications. As a rule of thumb: approximately 0.6 GB of VRAM per billion parameters with 4-bit quantization, plus a margin for context windows and system overhead.

Guideline values for Q4_K_M quantization with a moderate context window (approx. 8K tokens). Bar length = VRAM requirements relative to one another.

For comparison: A single consumer graphics card is generally not sufficient for a 70B model at full precision—this requires 80 GB of VRAM or more, which is only available on professional-grade cards. With 4-bit quantization, however, a 70B model can already be used with two RTX 4090 or RTX 5090 cards (totaling 48–64 GB) or a single workstation GPU. If the VRAM isn’t quite enough, tools like Ollama automatically offload some of the model layers to RAM—this works, but it noticeably slows things down.

💡 A practical introduction:

For mostbusiness scenarios in small and medium-sized enterprises, a single graphics card with 16–24 GB of VRAM (e.g., RTX 4090 or RTX 5070 Ti/5080), at least 32 GB of system RAM, and a power supply of 650 watts or more is sufficient. Ollama and LM Studio have established themselves as the go-to software platforms—both offer an OpenAI-compatible interface and can be set up in just a few minutes, without requiring a cloud connection. 

What scenarios can I use this for?

A locally operated 7B to 22B model is not a substitute for a Frontier model in the cloud—but it’s just right for a whole range of tasks, especially where data must not leave the premises:

🔒 Internal Document Assistant (RAG): Answer questions about internal manuals, quality management documentation, or technical specifications without the content everreaching anexternal server.

🧩 Legacy Code Analysis: A local code model like Codestral can review and document old source code in advance, before any sensitive business logic is even sent to a cloud tool—an approach we at VIALUTIONS use ourselves in our reverse-engineering practice.

🏭 Air-gapped environments: Production or GxP environments without an internet connection, where a model must run completely offline on local hardware to meet compliance requirements.

🌍 Multilingual design: Translations or text drafts in multiple EU languages, for which models such as Teuken-7B or EuroLLM-22B are inherently better suited than models trained exclusively in English.

🧪 Prototyping & Proof of Concept: Test and validate AI use cases before even committing to a paid cloud API connection.

💡 Realistically assess limitations:

Local 7B–22B models excel at clearly defined, well-documented tasks. For complex, multi-step reasoning, very long contexts, or tasks that require deep domain knowledge, Cloud Frontier models remain superior. A hybrid strategy—sensitive preprocessing on-premises, with complex analysis optionally performed in a sovereign cloud—is the most pragmatic approach for most companies. 

Digital Sovereignty “Made in Germany”: Three Approaches, Three Providers

Digital sovereignty is not a one-size-fits-all product—there are several valid approaches, and each of the three German providers mentioned represents its own distinct path. It is important to note: Sovereignty can refer to the model itself (open weights, self-hostable) or to the infrastructure used to process any given model (data center location, data agreements, GDPR compliance)—both are legitimate but distinct answers to the same question.

TextCortex (Berlin) AI Infrastructure

Positions itself as an enterprise AI infrastructure platform focused on secure, managed AI agents based on a company’s own data. Key components include its own RAG infrastructure, a zero-data training policy, and comprehensive security controls designed for DAX-listed corporations and German small and medium-sized enterprises. TextCortex does not train its own foundation model but rather builds governance and data protection frameworks around existing models.

 → textcortex.com/de 

Meloius (Southwestern Germany) Souvereign First

A platform launched in 2025 by four founders that consistently relies on European open-weight models: an API compatible with OpenAI and Anthropic interfaces for over 60 open models, along with vector storage and a chat interface. Operated exclusively on infrastructure based in the EU. Its core values: sovereignty, privacy by design, transparency, and sustainability from the very beginning.

 → melious.ai/de 

9brains (Bavaria) Confident Debut

Consolidates access to the world’s leading models from Anthropic, OpenAI, and Google through a single GDPR-compliant interface with a data center in Germany—with the guarantee that customer data will not be used to train external models. Complemented by agents that integrate with Microsoft 365, Brevo, Lexware, HubSpot, or your ERP system. “Sovereignty” here refers primarily to hosting and data agreements, not to a proprietary European foundation model. 

 → 9brains.de/ 

💡 The principle: 

“Digital sovereignty” is not a yes/no criterion, but rather a spectrum: ranging from purely GDPR-compliant data storage to a fully European, self-hostable model. Which point on this spectrum is the right one depends on the sensitivity of the data in each specific use case—not every request requires the highest level of sovereignty. 

What other European AI hubs are there?

In addition to individual model providers, Europe is currently working in parallel to build the necessary infrastructure—largely publicly funded and coordinated by the EU.

🏗️ EuroHPC Joint Undertaking (EU): This central European supercomputing initiative now coordinates 19 so-called AI Factories and several associated “Antennas” in member states ranging from Finland to Spain—with investments totaling around 10 billion euros for the period 2021–2027. In Germany, this includes the Jupiter AI Factory at the Jülich Research Center; in France, it includes AI2F, centered around the Jean Zay, Adastra, and Joliot-Curie supercomputers.

🏭 AI Gigafactories: A regulation that took effect in January 2026 expanded the EuroHPC mandate to coordinate the construction and operation of up to five even larger “AI Gigafactories” in cooperation with member states and the private sector — Calls for proposals have been open since the first quarter of 2026.

🔗 GAIA-X: A European initiative for federated, transparent data infrastructure—not an AI model in itself, but a set of rules and technical standards for sovereign, cross-border data exchange between European cloud and data providers.

🎓 National Data Centers & Research Clusters: The Barcelona Supercomputing Center (training ALIA), the Swiss Alps supercomputer in Lugano (training Apertus), and the Spanish MareNostrum 5 (training EuroLLM) demonstrate how academic high-performance computing centers have effectively become AI hubs.

🤝 OpenEuroLLM & Similar Consortia: Beyond individual models, lasting, EU-funded research consortia are emerging thatare continuouslyworking oncompetitive open language models for the European market.

💡 Context: 

“Most of these infrastructure initiatives are geared toward research, training new models, and supporting startups—rather than direct access to end customers. For companies looking for a ready-to-use product today, Mistral, Aleph Alpha/PhariaAI, and specialized platforms like the three mentioned above remain the more practical starting point.” 

Conclusion: Europe is no longer a niche market, but it’s not going it alone either

Anyone developing an AI strategy today doesn’t have to choose between “U.S. market leaders” and “nothing.” Between Mistral’s production-ready models, publicly funded open-source alternatives like Teuken or EuroLLM, and independent infrastructure providers such as TextCortex, Melious, or 9brains, a genuine European spectrum has emerged—ranging from cloud APIs to models running on one’s own graphics card.

The right choice rarely depends on model quality alone, but rather on how sensitive the processed data is, what compliance requirements apply, and how much control a company wants to retain over its infrastructure. It is precisely this assessment—which model, which hosting option, and which hardware for which use case—that is the first step in any sound AI strategy.

Which AI strategy is right for your systems?

VIALUTIONS supports companies in addressing legacy systems, data quality, and AI readiness, including the choice between cloud, hybrid, and on-premises models. 

→ Request a consultation