The Fable 5 export ban changed the conversation around open models in Europe. Before June 12, 2026, European companies could treat open-source AI as a nice-to-have. A cost optimization play. A hobby for tinkerers. After Anthropic was forced to disable its frontier models worldwide, open models became a strategic necessity.

But which one should you actually use? The three most relevant options for European organizations are Apertus (Swiss), Llama 4 (Meta/American), and Mistral Large 3 (French). They’re all “open” in some sense, but that word means very different things depending on who’s saying it.

I’ve spent time with all three. Here’s an honest comparison.

Quick comparison table

Feature	Apertus	Llama 4	Mistral Large 3
Origin	Switzerland (EPFL + ETH Zurich)	USA (Meta)	France (Mistral AI)
Largest model	70B dense	402B MoE (17B active)	675B MoE (41B active)
Smallest model	0.5B	109B MoE (17B active)	~3B (Mistral Small)
License	Apache 2.0	Llama Community License	Apache 2.0
Training data open	Yes, fully	No	No
Languages	1,800+	~8 primary	~12 primary
EU AI Act docs	Yes, published	No	Partial
GDPR compliance	By design	Unclear	Partial
Context window	4,096 tokens	10M tokens (Scout)	256K tokens
Multimodal	Text only	Text + Image	Text + Image
Training tokens	15T	Undisclosed	Undisclosed
Export control risk	None	Medium (US origin)	Low (EU origin)

Architecture and raw capability

Let’s start with what matters most to many developers: how good is the output?

Mistral Large 3

Mistral Large 3 is the most capable open model available today by most benchmarks. It’s a 675B parameter Mixture-of-Experts model with 41B active parameters, trained on NVIDIA H200 clusters. It launched as #2 among open non-reasoning models on the LMArena leaderboard. It handles coding, reasoning, and multilingual tasks at a level that genuinely competes with closed models.

It supports a 256K token context window and includes image understanding. The Apache 2.0 license means you can do whatever you want with it commercially. For raw capability, Mistral Large 3 is the open model to beat.

Llama 4

Llama 4 comes in two variants: Scout (109B total, 16 experts, 17B active) and Maverick (402B total, 128 experts, 17B active). Both are multimodal, handling text and images. Scout offers an unprecedented 10 million token context window.

The architecture is clever. By using MoE, Meta keeps the active parameter count at 17B while giving the model access to far more knowledge. Maverick with 128 experts can route to specialized knowledge efficiently. On English-language benchmarks, both models perform strongly.

But the license is a problem. The Llama Community License isn’t Apache 2.0. It includes restrictions that kick in above certain revenue thresholds, and it requires you to identify outputs as AI-generated in some contexts. More importantly, it’s a US-origin model governed by US law. If export controls tighten further (which seems likely given the trajectory), your Llama deployment could theoretically be affected.

Apertus

Apertus 70B is a dense transformer. No mixture of experts, no routing overhead, no architectural complexity. It was trained on 15 trillion tokens using a staged curriculum, with the xIELU activation function and AdEMAMix optimizer.

On English benchmarks, Apertus 70B performs in the range you’d expect from a 70B dense model. It’s solid. It’s competitive with Llama 3.1 70B era models. But it’s not in the same league as Mistral Large 3 or Llama 4 Maverick on English reasoning or coding tasks. That’s just the reality of dense 70B vs 675B MoE or 402B MoE.

Where Apertus shines is different. And that’s the point.

Multilingual performance: Apertus wins decisively

This isn’t close. Apertus natively supports over 1,800 languages. That’s not a typo. The training data was deliberately curated with 40% non-English content, covering languages that other models barely acknowledge exist.

Llama 4 primarily targets about 8 languages well (English, Spanish, French, German, Portuguese, Italian, Hindi, Thai). Mistral Large 3 covers roughly 12 languages with strong performance, focused on European and major Asian languages.

For European organizations working across EU member states, African development organizations, or anyone needing support for lower-resource languages, Apertus is simply the best option available. A Finnish government office, a Romanian hospital, or a Catalan media company will get better native-language performance from Apertus than from either competitor.

I’ve tested this with German, French, Italian, Dutch, and Portuguese prompts. Apertus handles them naturally without the subtle “translationese” quality you sometimes get from models that are primarily English-trained. The Swiss AI team’s decision to allocate 40% of training to non-English data pays off in practice.

Licensing and legal clarity

This is where things get interesting for enterprise deployments.

Apertus: Apache 2.0, clean and simple

No restrictions. No revenue caps. No usage limitations. You can deploy it, modify it, sell products built on it, and never tell anyone you’re using it. The Apache 2.0 license is as permissive as it gets.

Mistral Large 3: Also Apache 2.0

Mistral has committed to Apache 2.0 for their Mistral 3 series. This is a meaningful shift from their earlier models which had more restrictive licenses. For commercial use, Mistral Large 3 is equally permissive as Apertus from a licensing perspective.

Llama 4: It’s complicated

The Llama Community License looks open but has conditions. It requires attribution in certain contexts. It has provisions around monthly active user thresholds (700M+) that require separate licensing. And there are acceptable use restrictions that go beyond what Apache 2.0 imposes.

Is this a problem in practice? For most companies, probably not. You’re unlikely to hit 700M monthly active users. But the license complexity itself creates legal review overhead that Apache 2.0 simply doesn’t have. And the US-origin concern is separate from the license text.

Training data transparency

This is the differentiator where Apertus stands alone.

Apertus: Publishes everything. Training data reconstruction scripts are on GitHub. You can inspect exactly what went into the model, verify GDPR compliance yourself, respond to data subject access requests, and understand where any given output might originate from. They even publish EU AI Act transparency documentation and a Code of Practice.

Mistral Large 3: Provides some information about training methodology but doesn’t publish training data. You can’t independently verify what went in or respond to specific data subject requests about the training set.

Llama 4: Meta publishes a model card but the training data is completely opaque. You have no visibility into what data was used, whether it respects European data subjects’ rights, or whether it contains content that was obtained in ways that violate GDPR.

For any organization subject to regulatory scrutiny, operating in healthcare, finance, government, or any sector where you need to explain your AI systems to regulators, Apertus’s full transparency is enormously valuable. You can point a regulator at the training data and say “here, look at it yourself.” You can’t do that with Llama or Mistral.

EU compliance and the sovereignty question

After the Fable 5 ban, EU compliance isn’t just about GDPR anymore. It’s about operational continuity.

Supply chain risk

Apertus has zero US supply chain dependency. It’s developed by Swiss institutions, trained on Swiss supercomputers, and hosted on European infrastructure. Even if US-EU relations deteriorate further, your Apertus deployment keeps running.

Mistral Large 3 is nearly as safe. It’s developed by a French company, and France isn’t about to impose export controls on its own AI models going to EU customers. The risk here is if Mistral takes US investment that comes with strings attached, but so far they’ve maintained independence.

Llama 4 is the highest risk. Meta is a US company. The Llama license is governed by US law. If the export control regime expands (and the Fable 5 precedent suggests it might), there’s a non-zero chance that Meta could be instructed to restrict Llama access. The weights are already downloaded? Probably fine legally, since the license was granted. But future versions could come with new restrictions.

GDPR Article 17 (right to erasure)

If a European data subject asks you to delete their data from your AI system, you need to know what data is in there. With Apertus, you can check. With Llama and Mistral, you can’t. This is a real compliance gap that many organizations haven’t thought through yet.

Apertus even provides dedicated email addresses for PII removal and copyright requests: llm-privacy-requests@swiss-ai.org and llm-copyright-requests@swiss-ai.org. That’s institutional accountability.

Coding and reasoning capabilities

Let’s be honest about where Apertus falls short.

For coding tasks, the hierarchy is clear: Mistral Large 3 > Llama 4 Maverick > Apertus 70B. Mistral Large 3 with its 675B parameters and specialized training handles complex code generation, debugging, and refactoring at a high level. Llama 4 Maverick is strong for code as well. Apertus 70B can write code, but it’s not in the same tier for complex multi-file tasks or nuanced debugging.

For reasoning, similar story. Mistral and Llama have invested heavily in reasoning capabilities. Apertus can reason through problems, but it makes more errors on complex multi-step tasks.

If your primary use case is code generation or complex reasoning in English, don’t pick Apertus. Pick Mistral Large 3 if you want to stay in the EU ecosystem, or Llama 4 if you’re comfortable with the US-origin risk.

But if your use case is multilingual content, regulatory compliance, translation, summarization across European languages, or anything where data provenance matters more than peak performance, Apertus is the right choice.

When to use which model

Here’s my practical recommendation:

Choose Apertus when:

• You’re in a regulated industry (healthcare, finance, government)
• Multilingual support across many languages is critical
• You need to demonstrate training data compliance to regulators
• You want zero geopolitical supply chain risk
• You’re fine-tuning for a specific European-language task
• You need edge deployment (0.5B and 4B models)

Choose Mistral Large 3 when:

• You need the highest open-model performance on English tasks
• Coding and complex reasoning are primary use cases
• You want Apache 2.0 licensing with EU origin
• You have the hardware for a 675B parameter model
• Context window length matters (256K tokens)

Choose Llama 4 when:

• You need the 10M token context window (Scout)
• Multimodal (text + image) is important
• You’re already in the Meta ecosystem
• You accept the US-origin risk for better performance
• You need many specialized expert domains (Maverick’s 128 experts)

The bigger picture

None of these models replace each other perfectly. In a well-designed AI stack for a European organization, you might use all three:

• Apertus for compliance-sensitive tasks, multilingual content, and as your guaranteed-available fallback
• Mistral Large 3 for English-heavy tasks requiring maximum capability
• Llama 4 Scout for long-context document processing

The Fable 5 ban taught us that dependence on a single provider (or a single country’s models) is a business risk. Diversification isn’t just nice to have anymore. It’s prudent engineering.

Apertus isn’t the most powerful model in this comparison. But it’s the one that nobody can take away from you. In 2026, that matters more than most of us expected it would.

FAQ

Is Apertus actually competitive with Llama 4 and Mistral Large 3?

On English benchmarks and coding tasks, no. It’s in a different tier. On multilingual tasks across European languages, it’s genuinely the best. On compliance and transparency, it’s in a league of its own. The right answer depends entirely on your use case.

Can I switch between these models easily?

Yes. All three are supported in HuggingFace Transformers and vLLM. They all expose OpenAI-compatible APIs when served through vLLM. Your application code doesn’t need to change if you swap the model endpoint.

Which one should a European startup choose?

Start with Mistral Large 3 for maximum capability on your core product features. Add Apertus as a fallback and for any multilingual or compliance-sensitive features. Skip Llama unless you specifically need its context window or multimodal capabilities and are comfortable with the licensing situation.

Is the Fable 5 ban likely to affect Llama too?

The situations are different. Fable 5 is a closed API model. Llama’s weights are already distributed under an irrevocable license grant. But future Llama versions could come with new restrictions, and Meta could be pressured to stop releasing new weights internationally. It’s a low-probability but non-zero risk.

Does Mistral Large 3 really need 675B parameters worth of hardware?

No, thanks to the MoE architecture. Only 41B parameters are active per forward pass. You can run it on 8x A100 GPUs or even less with quantization. It’s resource-intensive but not as extreme as the total parameter count suggests.

What about fine-tuning? Which is easiest to customize?

Apertus is the easiest to fine-tune because it’s a standard dense transformer and everything about its training is documented. The 4B and 8B sizes are practical for fine-tuning on a single GPU. Mistral Large 3’s MoE architecture makes fine-tuning more complex. Llama 4’s MoE is similarly complex, though the community has developed good tooling around it.