If you’re a developer in Europe trying to figure out which AI models you can actually trust and deploy without geopolitical risk, the landscape has never been more confusing. Or more promising.

After the Fable 5 export ban showed that US AI access can disappear overnight, suddenly everyone’s talking about European sovereignty. But the truth is messy. Some projects are available today. Some are vaporware. Some are open. Some are commercial. Some are frontier-competitive. Most aren’t.

This guide maps every significant European sovereign AI effort in 2026. What exists, what’s coming, what you can use right now, and what’s just a press release with a timeline.

The landscape at a glance

Here’s the honest summary before we dive in:

Project	Origin	Status	Size	License	Frontier?
Apertus	Switzerland	Available now	8B / 70B	Apache 2.0	No
Mistral (various)	France	Available now	7B to Large	Mixed	Partially
OpenEuroLLM	EU consortium	Models due July 2026	TBD	Open	No
EUROPA	EU-funded	12-18 months	400B+ MoE	Open (TBD)	Goal
EuLLM	Various EU	Some available	7B range	Varies	No

Let’s break each one down.

Apertus: the best open European model you can use today

If I had to pick one model that represents European sovereign AI done right, it’s Apertus. Built by Switzerland’s top research institutions (EPFL, ETH Zurich, and CSCS) with Swisscom contributing, it ships as 8B and 70B parameter models under Apache 2.0.

What makes Apertus special:

• Trained on 15 trillion tokens across 1,811 languages (over 40% non-English data)
• Apache 2.0 license: Use it commercially, fine-tune it, deploy it however you want
• Fully documented: Training data, methodology, everything is transparent
• Available on Hugging Face right now: You can download and run it today

The 70B model is competitive with other models at the same scale. It handles European languages significantly better than most US-trained models, especially smaller ones like Swiss German, Romansh, and various regional variants. For multilingual applications serving European users, it’s genuinely excellent.

The limitation is straightforward: 70B isn’t frontier. You won’t get GPT-5.5 or GLM-5.2 level performance on complex reasoning, long-horizon coding tasks, or agentic workflows. It’s a solid workhorse, not a top performer.

For practical setup instructions, our guide to running Apertus locally walks you through deployment options.

Mistral: Europe’s commercial AI champion

Mistral is in a strange position. It’s the only European AI company with commercial-grade frontier models, hitting $400M in annual recurring revenue in early 2026 and raising a $2B Series C at a $13.8B valuation. But “European” here comes with asterisks.

Mistral’s models range from the open-weight Mistral 7B (Apache 2.0) to their larger commercial models available through la Plateforme and Le Chat. Since the Fable 5 ban, they’ve leaned heavily into the sovereignty narrative. As one TNW article put it, “the Anthropic shutdown has handed Mistral its sovereignty argument on a plate.”

What Mistral offers developers:

• Open-weight smaller models (7B) for self-hosting
• Commercial API access to larger models
• Enterprise deployment options (“Build Your Own AI” strategy)
• European data residency guarantees
• Le Chat as a consumer-facing product

The caveats:

• Larger models aren’t fully open-weight. You can’t download and run them yourself.
• Pricing is commercial. This isn’t free infrastructure.
• They’ve taken investment from non-European sources, which complicates the “pure sovereignty” story.
• The focus is commercial customers, not open research.

For a detailed comparison of how Mistral stacks up, see our Apertus vs Llama vs Mistral comparison.

Mistral is great if you want enterprise-grade AI with European data handling guarantees and you’re willing to pay for it. It’s not what you want if you need fully open weights you control end-to-end.

OpenEuroLLM: the pan-European research effort

OpenEuroLLM is a consortium of Europe’s leading AI companies and research institutions collaborating to build open-source multilingual language models. It’s funded through EU programs and includes contributions from organizations across multiple countries.

The project has three key goals:

1. Build the first family of open-source LLMs covering all official EU languages
2. Develop infrastructure for large-scale distributed training across European clusters
3. Create evaluation frameworks for multilingual AI performance

Current status (June 2026): One year into a three-year project. First models are expected in July 2026. According to multiple reports, the project faces “significant resource constraints, primarily in compute capacity.” This is the recurring challenge for European AI: we have the talent and the ideas, but not always the GPUs.

OpenEuroLLM is important for the ecosystem because it’s building shared infrastructure, not just one model. The training pipelines, evaluation tools, and multilingual datasets they create will benefit all European AI projects, including EUROPA.

But let’s be real: if you need a model in production today, OpenEuroLLM can’t help you yet. Check back in July 2026 for their first release.

EUROPA: the frontier bet

The newest and most ambitious entry. The EU Commission selected the EUROPA consortium (led by Italian company Domyn) in June 2026 to build a 400B+ parameter model with open weights covering all 24 EU languages. It’s funded through the Frontier AI Grand Challenge and will run on EuroHPC infrastructure.

EUROPA is designed to be the thing Europe has never had: an open model that can compete with the best in the world. Not “good for a European model.” Actually frontier-competitive.

The timeline is 12-18 months. That means late 2027 at the earliest. For the full breakdown of what EUROPA is and what it means, see our detailed EUROPA article.

Why EUROPA matters for the ecosystem: It validates that the EU is willing to fund frontier-scale compute. Previous European AI efforts have been limited by compute budgets. EUROPA, backed by EuroHPC’s supercomputers, shouldn’t have that constraint.

EuLLM and national initiatives: the long tail

Beyond the headline projects, dozens of smaller efforts are training models for specific languages and use cases:

• Various national language models at the 7B parameter range for languages like Finnish, Estonian, Latvian, and others that get minimal attention from US-trained models
• Domain-specific models for healthcare, legal, and public administration use cases
• Fine-tuned versions of Apertus and other base models for national deployment

These won’t make headlines, but they matter. A 7B model fine-tuned specifically for Estonian legal text might outperform a 400B general model for that specific task. Sovereign AI isn’t just about frontier capability. It’s about having models that understand your specific context.

What can you actually use RIGHT NOW?

Let’s cut through the announcements and get practical. If you need European-hosted, sovereign AI today, here are your real options:

For open self-hosted deployment:

1. Apertus 70B - Best overall open European model. Apache 2.0. Strong multilingual performance. Get started here.

2. Apertus 8B - Same architecture, lighter weight. Good for edge deployment or when you need fast inference on modest hardware.

3. Mistral 7B - Open weights, Apache 2.0. Good general performance but less multilingual focus than Apertus.

For commercial API access:

1. Mistral’s la Plateforme - European-hosted, GDPR-compliant API. Multiple model sizes. Commercial pricing.

2. Various European cloud providers hosting open models (Apertus, Mistral 7B) on EU infrastructure.

For frontier capability (with caveats):

Here’s the uncomfortable truth: if you need frontier-level AI performance today and you’re restricted to European sovereignty, you don’t have a fully sovereign option. Your choices are:

• Chinese open models like GLM-5.2 (MIT license, 744B MoE) or DeepSeek V4. Frontier-capable and fully open, but training data and organizational ties create data sovereignty questions for some use cases.
• Mistral’s commercial models via API. European company, but not open-weight at the frontier tier.
• Wait for EUROPA (12-18 months).

This gap is exactly what EUROPA is meant to fill. But today, you have to make trade-offs.

The compute problem

Every article about European AI eventually hits the same wall: compute. Training a frontier model requires thousands of high-end GPUs running for months. Europe has historically under-invested in this infrastructure compared to US and Chinese hyperscalers.

EuroHPC is supposed to fix this. The EU’s joint supercomputing initiative operates some of the world’s most powerful machines. The EUROPA project will train on these resources. OpenEuroLLM has struggled with compute constraints.

The math is simple. You can have the best researchers in the world (Europe does), the best data curation (Europe arguably does for multilingual), and the best regulatory framework (debatable, but the AI Act exists). Without GPUs, none of it becomes a model.

The Frontier AI Grand Challenge represents a commitment to solve this. Whether the execution matches the ambition remains the open question.

How the Fable 5 ban changed everything

It’s worth emphasizing how much June 12, 2026 changed the calculus for European businesses. Before the Fable 5 export ban:

• “Sovereign AI” was a nice-to-have, a political talking point
• Most European enterprises happily used OpenAI, Anthropic, or Google APIs
• The sovereignty risk was theoretical

After the ban:

• Enterprises lost access to their primary AI tools overnight
• Legal teams started asking about AI supply chain risk
• Procurement policies changed. Vendors now need to prove they can’t be cut off.
• European alternatives went from “maybe later” to “urgent requirement”

Mistral saw this immediately. Their sovereignty pitch, previously seen as nationalistic marketing, suddenly became the most compelling enterprise argument in AI. The company’s CEO sat down with Anthropic’s and OpenAI’s CEOs at the G7 summit just days after the ban, from a position of unprecedented strength.

For developers, the lesson is practical. If you’re building anything critical, you need at least one fallback that can’t be export-controlled. Today, that means open-weight models hosted on infrastructure you control.

What’s coming in the next 12 months

Here’s the timeline European developers should watch:

• July 2026: OpenEuroLLM first model release
• Late 2026: Likely updates on EUROPA training progress
• Early 2027: Possible intermediate EUROPA checkpoints or smaller models
• Mid-to-late 2027: EUROPA full model release (optimistic estimate)
• Ongoing: Mistral continues releasing new commercial models
• Ongoing: National initiatives releasing specialized models

The landscape will look very different a year from now. But you don’t have to wait. The tools available today, particularly Apertus and Mistral, are production-ready for most use cases that don’t require bleeding-edge frontier capability.

My take: what developers should actually do

Stop waiting for the perfect European model. Here’s my practical recommendation:

1. Deploy Apertus 70B today for any workload where you need sovereignty guarantees and the 70B scale is sufficient. That covers most production use cases.

2. Use Mistral’s API for cases where you need more capability but can accept commercial terms and API dependency (still European, still GDPR-compliant).

3. Evaluate GLM-5.2 and DeepSeek Vision for frontier capability if your use case allows Chinese-originated models with MIT licenses. The weights are open and you can self-host them on European infrastructure.

4. Watch EUROPA and plan your architecture so you can swap in the frontier European model when it arrives. If you build on open standards and model-agnostic APIs today, the migration will be trivial.

5. Don’t put all eggs in one basket. The Fable 5 ban proved that any single dependency is a risk. Build with model-switching capability from day one.

European sovereign AI in 2026 is no longer a dream. It’s just not yet complete. The foundation exists. The frontier is coming. And for the first time, there’s genuine urgency and funding to make it happen.

FAQ

What’s the best European AI model I can use today?

For open-weight self-hosting, Apertus 70B is the strongest option. For commercial API access with European data residency, Mistral’s models are the most capable. Neither is frontier-competitive with GPT-5.5 or GLM-5.2, but both are production-ready.

Is Switzerland considered part of European sovereign AI?

Switzerland isn’t in the EU, but Apertus is open-source under Apache 2.0, so it’s available to everyone. The Swiss institutions (EPFL, ETH Zurich) collaborate closely with EU research networks. For practical purposes, Apertus is part of the European AI ecosystem even if it’s not EU-funded.

When will Europe have a frontier-competitive open model?

The EUROPA project targets delivery in 12-18 months (late 2027 to early 2028). OpenEuroLLM has models coming in July 2026, but they aren’t expected to be frontier-scale. Mistral has commercial frontier models now, but they’re not fully open-weight.

Can I use Chinese open models on European infrastructure?

Yes. Models like GLM-5.2 (MIT license) and DeepSeek V4 (MIT license) can legally be downloaded and deployed on any infrastructure, including European servers. The weights are open. The concern isn’t legal but rather supply chain: do you trust the training data and model provenance for your specific use case?

What does the EU AI Act mean for model deployment?

The AI Act’s requirements for general-purpose AI models (transparency, documentation, copyright compliance) apply progressively through 2027-2028. If you deploy open models, you’ll need to ensure compliance with documentation and transparency requirements. EUROPA is being designed AI Act-compliant from the start, which could make it the easiest model to deploy in regulated European contexts.

How much does it cost to run Apertus locally?

The 8B model runs on consumer hardware (a single high-end GPU). The 70B model requires more serious infrastructure: multiple GPUs or a dedicated inference server. Cloud deployment on European providers is the practical middle ground for most teams. See our complete local deployment guide for hardware requirements.