Switzerland's Apertus is fully open down to the data — and has no GPUs
EPFL, ETH Zurich and the national supercomputing center shipped a foundation model where the training data, code, weights and alignment are all published. The unusual part: the license warns its output may contain personal data, and ships you a hash file of deletion requests to filter it yourself.
Apertus ships in 8B and 70B sizes, claims more than 1,000 languages, and bakes in EU-AI-Act compliance: PII removal, opt-out respect, memorization prevention. The pitch is total reproducibility — "to AI as Open is to Source" — not benchmarks. Cutoff: March 2024.
One founder reports building real RAG software on it today; it is not ready for agentic work, but it drives. Skeptics aimed at the institution: a model moving at committee speed, a country with no domestic GPUs, and a project whose best output may be the people it trained.
Open all the way down to the training set
Apertus is the Swiss AI Initiative's bid for what it calls sovereign AI, built by EPFL, ETH Zurich and the CSCS supercomputing center and released in 8B and 70B sizes. The differentiator is not the size; it is the disclosure. Training data, code, weights, methods and alignment are all documented and reproducible, a stance the project frames as being "to AI as Open is to Source." It claims to be competitive with top open models at equivalent scale, spans more than 1,000 languages, and carries a March 2024 knowledge cutoff.
Compliance written into the license
The EU-AI-Act framing is more than a slogan here: PII removal, respect for opt-outs, and memorization prevention are listed as design goals, with recent additions including an Apertus Mini in June 2026 and an ACL 2026 paper. The license then makes a move most model cards avoid. It warns that output may contain personal data, and says SNAI will periodically publish a file of hash values reflecting deletion requests, which users are expected to download and apply as an output filter on their own usage. The release drew 236 points and 83 comments, and the thread was where ambition met the ledger.
A working model versus a committee's pace
The most grounded endorsement came from someone shipping with it.
"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." — atemerev, founder at Virial and ex-EPFL
The doubt was structural, aimed at institutional velocity rather than the weights.
"the impression I get from Apertus is that it moves at the speed of a committee. I have no expectation they'll deliver a competitive model." — a YC W07 founder who works on Virtualmin
A blunter constraint cut underneath the sovereignty pitch entirely.
"the swiss have no gpus" — nisten, an open-source AI researcher with many models and datasets on Hugging Face
And one reader reframed the mediocrity as a down payment.
"the most impactful product of the Apertus project are the people… I bet if the team would do another model training they could get way better results at one fourth of the costs." — a commenter quoting Dominique Paul
Sovereignty you can audit, not yet sovereignty you can scale
The contradiction sits in the name. Apertus delivers genuine sovereignty over the artifact — you can inspect every layer of how it was made, which no frontier lab offers — yet sovereign AI also means controlling the compute, and a model trained in a country with no domestic GPUs at the speed of a committee is sovereign in provenance more than in power. If the project's real yield is a trained team that could halve the cost next time, then the first Apertus is best read as tuition, not as the product.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Did my old job only exist because of fraud? A decade later, the SEC answers
A software engineer's first job carried him from the UK to the US, where he built a life and met his spouse. Years on, regulators alleged the acquirer ran an incubator mostly to harvest management fees.