Apertus: Europe's Answer to AI Sovereignty - and Why HN Is Skeptical

The Swiss AI Initiative just shipped Apertus, a fully open foundation model developed by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS). The pitch: complete transparency, EU AI Act compliance, and training data you can actually inspect.

With 497 points and 167 comments on Hacker News, the response is... complicated.

What Apertus Actually Is

Apertus ships in two sizes - 8B and 70B parameters - trained on 15 trillion tokens across 1,000+ languages. The headline differentiator is openness: training data, code, weights, methods, and alignment principles are all documented and reproducible.

From the official site:

Fully open model: open weights + open data + full training details including all data and training recipes.

The model was trained on CSCS's Alps supercomputer using up to 4,096 GPUs. Swisscom serves as the strategic partner, and you can access it through Hugging Face or the Public AI network.

Key claims:

• Built to meet EU AI Act requirements
• Respects opt-outs and removes PII
• Prevents memorization
• Competitive with top open models at equivalent scales

What HN Is Saying

The HN discussion splits into a few camps:

The "Actually Open" Appreciation

The core appeal is scientific reproducibility. One commenter put it well:

As long as the following remains true, this release ends up a bigger contribution to science at large than most other models trained "behind closed doors."

Several users pointed out that Apertus joins a small club of genuinely open models - alongside Allen AI's OLMo 3.1, MBZUAI's K2 Think V2, and Nvidia's Nemotron (though Nemotron has some proprietary data).

The Performance Question

The most upvoted criticism is simple: is the model actually good?

One user shared benchmarks from Artificial Analysis showing Apertus trailing Nemotron significantly. The comparison link shows Nvidia's Nemotron-3-Ultra-550B and Super-120B outperforming both Apertus variants on standard intelligence benchmarks.

Another commenter reported:

The previous version of this model has been pretty bad, but claimed to adhere to copyright laws. However, based on my testing, that's not true either.

The multilingual claims also drew skepticism. One user noted that for "simple questions like 'how to say X in language Y' or 'how to conjugate verb X in language Y'" - the model "keeps hallucinating words that do not exist."

The Training Data Debate

This is where things get philosophically interesting. Simon Willison noted that Apertus "uses fineweb, which is derived from Common Crawl, which is an unlicensed scrape of web pages."

This prompted a subthread about whether "sovereign AI" can really claim ethical high ground while using scraped data. One response:

You don't need a license to scrape the public web and analyze it, turn it into tokens and other transformations. Let's not expand copyright beyond the horrible monster it already is.

Willison's counterpoint:

That doesn't mean much to the many people I know of who refuse to use a technology that they see as being unethically created using the work of others without compensating them.

The Geopolitical Angle

Several comments touched on why European AI independence matters. One user argued:

Frankly, I'm surprised there's not more urgency on the part of Europeans to reduce dependence on US tech.

The argument: even if Apertus isn't state-of-the-art today, having European-controlled AI infrastructure matters for data sovereignty and regulatory compliance.

The Practical Takeaway

If you're evaluating Apertus for production use, here's what the evidence suggests:

Strengths:

• Full training transparency - you can audit everything
• EU AI Act compliance baked in
• No licensing restrictions for commercial use
• Multilingual training with underrepresented languages (Swiss German, Romansh)

Weaknesses:

• Performance lags behind Nemotron, DeepSeek, and frontier models
• Not ready for agentic use cases according to hands-on reports
• Multilingual quality is inconsistent in practice

Best use cases:

• RAG applications where transparency matters more than peak performance
• European organizations with strict data sovereignty requirements
• Research where reproducibility is the priority

One commenter who actually uses Apertus in production:

I use it extensively. It is not ready for agentic use, but as a generic driving model for RAG use cases, it is pretty competent. You can build useful software with it.

The Bigger Picture

The debate in this thread mirrors a broader tension in AI development: should we prioritize transparency and ethical sourcing even at the cost of capability? Or does "open" only matter if the model is actually competitive?

For now, Apertus represents a credible attempt at the transparency-first approach. Whether the capability gap closes depends on continued funding and research. The Swiss AI Initiative has the institutional backing - EPFL and ETH Zurich are serious players - but catching up to labs spending billions requires more than good intentions.

If you want a model where you can trace every training decision, Apertus delivers. If you need frontier performance, you're still looking at Nemotron, DeepSeek, or the closed models.

The market has room for both.

Sources

• Apertus Official Site
• Hacker News Discussion
• ETH Zurich Press Release
• Artificial Analysis Benchmarks