Switzerland now has its own open language model
EPFL, ETH Zurich and the national supercomputing centre released Apertus today. For once, a model release that is genuinely a Swiss story.
On 2 September 2025, EPFL, ETH Zurich and the Swiss National Supercomputing Centre released Apertus, a fully open multilingual language model, in two sizes: eight billion and seventy billion parameters. It was trained on the Alps supercomputer in Lugano.
This blog has covered a great many model releases without much excitement. This one is different, and it is worth explaining precisely why — and also what it does not mean.
What « fully open » means here
Most so-called open models publish the weights and nothing else. Apertus publishes the weights, the training data composition, and the recipe used to build it.
That is a meaningfully higher standard. It means an independent party can examine what went into the model rather than taking a provider’s word for it, which is the entire difference between transparency as a claim and transparency as a property.
The multilingual point, which matters here more than elsewhere
Roughly forty per cent of the training data is non-English, spanning over a thousand languages. This blog has repeatedly noted that most models handle German and Italian less well than English, and that the gap should shape what you let them do unreviewed.
A model built by a four-language country, deliberately, is a reasonable place to expect that gap to be smaller. Whether it actually is will be established over the coming months by people testing it, not by the announcement.
What this does not mean for your company
It does not mean you should switch to it. It is a research release from a public institution, not a product with a support contract, and the practical questions from 6 August apply unchanged: who runs it, on what, and who fixes it at four on a Friday.
What it does mean is that « there is no European alternative » is no longer an accurate thing to say, and that a Swiss company asking for a Swiss-built option now has an answer to point at.
Thursday’s post looks at the honest version of the follow-up question: could a company your size actually run this?