EPFL, ETH Zurich, and the national supercomputing centre published an 8B and 70B LLM with open weights, data, and training code — a public-interest answer to closed AI.
2026-06-22 · by Moe Ameen
On September 2, 2025, the Swiss AI Initiative — a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS) — released Apertus, the country's first large-scale, fully open multilingual language model. The name is Latin for "open," and that is the point: alongside the model weights, the team published the training data, source code, training recipe, and alignment principles, all documented and reproducible. That goes beyond the common "open weights" release, which ships a checkpoint without the data or method behind it.
Apertus comes in two sizes, an 8-billion-parameter model and a 70-billion-parameter model, each available as a pretrained base and an instruction-tuned variant under the permissive Apache 2.0 license, which allows commercial use. It was trained on roughly 15 trillion tokens spanning more than 1,000 languages, with about 40% of the data outside English — including languages usually left out of large models, such as Swiss German and Romansh — and supports a long 65,536-token context.
The team built the model with EU AI Act transparency obligations and Swiss copyright and data-protection law in mind: the corpus uses only publicly available data, filtered to respect machine-readable opt-outs (even retroactively) and to remove personal information. The institutions framed the release as a building block for "sovereign AI" — technology that public bodies, companies, and researchers can run and audit on their own infrastructure rather than renting from a closed provider. Apertus is available to download from Hugging Face and through partners including Swisscom and the Public AI Inference Utility. It is a text-and-reasoning model; it does not generate images, video, or audio.
The open-versus-closed and sovereign-AI debate is exactly the kind of high-intent topic your audience is reading about right now, and a clear take on it is content. Drop your point of view — why open models matter, who should care about data sovereignty, when self-hosting is worth it — into Kompozy as a source, and the engine fans that single take into a blog post, a carousel explainer, short captioned clips, and platform-native posts in your voice, then schedules and publishes across all nine connected platforms. Being early and specific on a story like this is how one opinion becomes a week of coordinated content.
There is a practical workflow too, especially for multilingual creators acting on this moment. You can draft or translate copy in Apertus — even in a self-hosted deployment if data has to stay in-house — and then bring those drafts into Kompozy to do everything the model cannot: render persona and avatar video, build carousels and quote cards in your brand style, caption and reframe clips per platform, and publish on a schedule. Apertus supplies the open, multilingual words; Kompozy turns them into finished, localized, on-brand posts across every channel.
Apertus is a fully open, multilingual large language model released on September 2, 2025 by the Swiss AI Initiative — a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS). It ships in 8B and 70B sizes and publishes its weights, training data, code, and methods openly under the Apache 2.0 license.
Its entire pipeline is open and reproducible — weights, training data, code, and alignment — and it was built around EU AI Act transparency and Swiss data-protection law. That lets governments, public institutions, and companies run and audit the model on their own infrastructure instead of depending on a closed foreign provider.
It can draft text, but it cannot generate images or video, design posts, or publish anything — it is a text-and-reasoning model with no media or distribution layer. To turn an Apertus draft into published content you pair it with a generation and publishing engine like Kompozy, which renders the media and publishes across nine platforms.
Yes. The model is released under Apache 2.0, which permits research, education, and commercial use at no license fee. You can download the weights from Hugging Face or access it through partners like Swisscom and the Public AI Inference Utility; your real cost is the infrastructure to run it.