Apertus
Open language model from Switzerland with published weights and training data, for self-hosting or use via Swiss hosting providers.
What it does
Apertus is an open language model from the Swiss AI Initiative, developed by EPFL, ETH Zurich and the Swiss National Supercomputing Centre CSCS. Version 1.5 comes in 8B and 70B sizes. Besides text it understands images and, still experimentally, audio, but it answers in text only. An optional thinking mode lets the model reason before answering, and the context window holds around 262,000 tokens. Weights, training data, code and methods are published; training covered text in more than 1,000 languages.
Who it suits
Companies and developer teams who want a transparent model they can run themselves or integrate via API from a Swiss host, for example for internal assistants. Guides exist for Ollama, LM Studio, Open WebUI and vLLM, among others. The Apache 2.0 licence permits commercial use.
Data protection for businesses
If you run Apertus on your own hardware, your inputs stay with you. Hosted versions are offered by Swisscom, Infomaniak and Phoeniqs, among others, with hosting in Switzerland. You settle the data processing agreement, any training use of your inputs and business plans with the respective host; the project itself sells no plans. According to the project, website opt-outs were respected during training and personal data was removed.
Limits
The project offers no chat product of its own; for trying it in the browser it points to third parties such as Public AI. Guides for local use are partly still being written. Audio processing is experimental, and performance comparisons come from the project itself.