The Swiss AI Initiative just dropped Apertus, a fully open foundation model that claims competitive performance at 8B and 70B scales while training on over 1000 languages. That multilingual breadth alone should make developers outside English-first markets sit up.
1000-Language Training and Open Weights
Apertus ships open weights, open data, open code, and open alignment principles. The project is a collaboration between EPFL, ETH Zurich, and CSCS, with Swisscom as a strategic partner. They're not vague about openness: "Apertus is to AI as Open is to Source." Apertus Mini, a set of 16 small language models released June 15, already demonstrates distillation and quantization techniques from the main model.
Built for EU AI Act Compliance
This is the first model I've seen that explicitly bakes EU AI Act requirements into the training pipeline. The team says Apertus respects opt-outs, removes PII, and prevents memorization. That's a concrete architectural decision, not a policy sticker. If you need a foundation model that can legally slot into European workflows without a rewrite, Apertus is the most credible candidate so far.
What Ships Now and What's Next
The technical report has been accepted at ACL 2026, meaning the community will get a peer-reviewed look at the methods. A fine-tuned variant called Apertus for Ticino already powers in-house AI translation. I'd be wary of marketing language, but the concrete numbers and EU AI Act compliance targets make this worth watching. Apertus positions itself as the sovereign AI foundation for Europe, and the ACL 2026 paper will put its reproducibility claims to peer review.
Source: Apertus - Open Foundation Model for Sovereign AI
Domain: apertvs.ai
Comments load interactively on the live page.