Switzerland has quietly built what every sovereign AI initiative has been promising: a fully open foundation model where the weights, training data, code, methods, and alignment principles are all documented and reproducible. Apertus, developed by researchers at ETH Zurich and EPFL at the Swiss National Supercomputing Centre (CSCS), comes in 8B and 70B parameter sizes under Apache 2.0, is trained on 1000+ languages, and is explicitly designed to meet EU AI Act requirements. It hit 167 points on Hacker News over the weekend — the community signal that this isn’t just another government press release.
🔍 THE BOTTOM LINE
Apertus is the model that proves sovereign AI doesn’t have to mean closed AI. While countries pour billions into national champions with restricted access, Switzerland built the opposite: a model where you can read the training data, audit the alignment, run it locally, and modify it. For New Zealand — which has been grappling with its own sovereign AI moment — this is the blueprint worth studying.
What “Fully Open” Actually Means Here
Most “open” models release weights and call it a day. Apertus goes further. The team has published their training data reconstruction scripts, the full codebase for training and evaluation, a custom chat format library, and alignment documentation. The model card on Hugging Face shows 70.6 billion parameters in BF16, 32,000+ downloads, and an accepted paper at ACL 2026 titled “Apertus: Democratizing Open and Compliant LLMs for Global Language Environments.”
The distinction matters because “open weights” and “open source” are not the same thing. As we’ve covered in our analysis of open models beating big tech, the gap between proprietary and open has narrowed to months, not years. Apertus closes a different gap: the one between “you can download the weights” and “you can reproduce the entire pipeline.”
Built for Compliance, Not Just Capability
Apertus was designed from inception to meet EU AI Act requirements. The model respects opt-outs, removes PII, prevents memorization, and ships with a Swiss AI Charter as its internal constitution. There’s a public EU summary for AI Act compliance and full documentation of the code of practice. This isn’t a model that was built first and then compliance-washed after the fact — the regulatory architecture is part of the foundation.
For a Swiss project, that positioning is natural. Switzerland sits inside the European regulatory orbit without being an EU member, giving it both the incentive to comply and the independence to innovate. Swisscom is a strategic partner, which means the model has a commercial deployment path from day one — not just academic prestige.
The Sovereign AI Race Has a New Entry Pattern
The sovereign AI conversation has been dominated by two approaches: build a closed national champion (France’s Mistral, China’s Baidu) or buy foreign and hope for the best. Switzerland is proposing a third way: build open, let anyone verify it, and compete on trust rather than lock-in. This is a meaningful contrast to Mistral’s €3 billion raise at a €20 billion valuation — both are European sovereign AI plays, but they represent fundamentally different theories of how to compete.
It also contrasts with the APAC sovereign AI race where NZ is being left behind. The APAC approach has been infrastructure-heavy: buy GPUs, build data centres, train models behind national firewalls. Switzerland’s approach is lighter and more replicable: use public supercomputing (CSCS), open the entire stack, and let the global community stress-test it. For a small economy with limited compute budget, that’s a more viable path.
NZ Angle
New Zealand’s sovereign AI conversation has focused on data residency and indigenous data sovereignty — important but incomplete. Apertus demonstrates that the model layer matters too. A NZ organisation could download Apertus today, fine-tune it on local data, and run it on local infrastructure with full knowledge of what went into the base model. That’s not possible with Claude, GPT, or even most “open” models that stop at weights.
The UK, Korea, and NZ sovereign AI framework comparison we published earlier identified the policy gap. Apertus fills the technical gap. The combination — a sovereign AI policy plus an actually-open foundation model — is what a credible national AI strategy looks like. NZ has neither right now, but the Swiss model shows the cost of building one is lower than expected when you start from an open base.
❓ FAQ
Can NZ businesses use Apertus commercially? Yes. Apache 2.0 is one of the most permissive licenses available. You can use it, modify it, sell products built on it, and distribute derivatives — all without royalties. The only requirement is preserving the copyright and license notice.
How does 70B compare to what’s available from OpenAI or Anthropic? At 70B parameters, Apertus is competitive with other open models at equivalent scale (Llama 3 70B, Mistral Large). It won’t match GPT-5 or Claude Opus on raw benchmarks, but for most enterprise use cases — classification, summarisation, retrieval, translation — the gap is marginal and shrinking. The EU AI Act compliance layer is something no US provider offers.
What’s the Apertus Mini collection? The team released 16 smaller models (0.5B, 1.5B, 4B) distilled from the full 8B model, designed for edge devices and compute-constrained environments. This matters for NZ: you can run a 4B model on a MacBook, no data centre required.
Is “trained on 1000+ languages” actually useful? For low-resource languages including te reo Māori, multilingual pre-training provides a stronger base than English-only models. The Apertus team has already demonstrated a fine-tuned model for Swiss Italian dialect — the same approach works for any language with sufficient text data.
🔍 THE BOTTOM LINE
Apertus is what sovereign AI was supposed to look like before every country decided to build its own walled garden. Switzerland bet on radical transparency, and the result is a model that’s more trustworthy, more compliant, and more replicable than anything the national-champion approach has produced. For NZ, the lesson isn’t to copy Switzerland — it’s to stop pretending sovereign AI means closed AI.
