Apertus
Open multilingual language models (8B and 70B) with published weights, data and code.
- Swiss CompanyThe operating company is headquartered and legally incorporated in Switzerland.
- Swiss OwnedThe ultimate controlling owner is Swiss (Swiss persons, companies, foundations or public bodies).
- Swiss DevelopedThe product is primarily developed by a team operating from Switzerland.
- Swiss Data OptionSwiss hosting is available as an option or for part of the service, but is not the default for everything.
- Open SourceThe product’s source code is published under an open-source licence (client and server).
- Self-hostableCan be installed on your own servers or on-premises.
- Free PlanA permanently free plan is available (not just a trial).
apertus-ai.org
Overview
Apertus is a family of open large language models developed by EPFL, ETH Zurich and CSCS as part of the Swiss AI Initiative. The first release in September 2025 comprised base and instruction-tuned models with 8 and 70 billion parameters, trained on 15 trillion tokens covering more than 1,000 languages, about 40% of them non-English, including Swiss German and Romansh. Apertus 1.5, with weights released in July 2026 and documented in October 2026, extends both sizes with image and audio input, an optional reasoning mode and a context window of up to 262,144 tokens.
The weights are published on Hugging Face under the Apache 2.0 licence, and the project also publishes its technical report, training code and the scripts used to reconstruct the training data. Apertus 1.0 was trained on the Alps supercomputer operated by CSCS in Lugano. Apertus is a set of models rather than a finished chat product; the Apertus 1.0 model card notes that no output filter is provided.
Apertus can be run on one’s own hardware or used through third-party APIs listed by the project, several of which are hosted in Swiss data centres (for example Swisscom, Infomaniak and PHOENIQS). Apertus 1.5 is also offered in the Proton Lumo and Infomaniak Euria assistants.
Why it may be a Swiss alternative
Apertus is operated by Swiss AI Initiative, which is headquartered in Lausanne, Vaud. Swiss-headquartered, Swiss-owned company; product developed (wholly or largely) in Switzerland.
Swiss hosting is available for part of the service or as an option — check the hosting details below before relying on it.
May replace
Foreign products with overlapping functionality. An alternative is not necessarily a one-to-one replacement.
Swiss footprint
Five independent questions. A Swiss company is not automatically Swiss-hosted — and vice versa. How the footprint is assessed
Swiss native. Swiss-headquartered, Swiss-owned company; product developed (wholly or largely) in Switzerland.
- HeadquartersIs the operating company headquartered in Switzerland?
- Lausanne, Vaud (VD), SwitzerlandInferred[2]
The initiative is not a separate legal entity; it is co-led by EPFL (Lausanne) and ETH Zurich (Zurich), with CSCS in Lugano providing compute. Lausanne is given because EPFL is named as the owner of the Apertus website; Zurich would be equally defensible.
- Legal jurisdiction
- SwitzerlandInferred[2]
- OwnershipIs the company ultimately Swiss-owned?
- Yes — ETH Domain (EPFL and ETH Zurich) · SwitzerlandVendor claim[1]
The Swiss AI Initiative is the first project of the Swiss National AI Institute, a partnership of the ETH AI Center and the EPFL AI Center. It was seeded with a CHF 20 million grant from the ETH Domain and more than 10 million GPU hours on the Alps supercomputer at CSCS. EPFL and ETH Zurich are federal institutes of technology.
- DevelopmentIs the product developed in Switzerland?
- YesVendor claim[1]
Developed by researchers at EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) within the Swiss AI Initiative.
- HostingIs customer data hosted in Switzerland?
- OptionalVendor claim[5]
Apertus is distributed as downloadable weights and runs wherever it is deployed. Hosted APIs from Swiss data centres are listed for Swisscom, Infomaniak, PHOENIQS, Safe Swiss Cloud, OnPrem.ai and stepping stone, with deployment guides for Exoscale; Public AI, Featherless and the large US cloud platforms also offer it. The infrastructure below refers to model training.
- InfrastructureIs the hosting infrastructure operated by a Swiss-owned provider?
- Yes — Swiss National Supercomputing Centre (CSCS) · SwitzerlandVendor claim[1]
Apertus 1.0 was trained with more than 10 million GPU hours on Alps, using up to 4,096 NVIDIA GH200 GPUs. Data-centre facility: Alps supercomputer, CSCS Lugano
Features & platforms
- Platforms
- Unknown / not verified
- Deployment
- Made for
- Interface languages
- Unknown / not verified
- Launched
- 2025
Privacy & security
- Not verified
- End-to-end encryption, Encryption at rest, Encryption in transit, Multi-factor authentication, Single sign-on (SAML/OIDC), SCIM provisioning, Public API and Data export
- Certifications
- None listed
Pricing
- Pricing model
- Free
- Free plan
- Yes
- Free trial
- Unknown / not verified
- Enterprise plan
- Unknown / not verified
- Billing in CHF
- Unknown / not verified
Prices checked 15 Oct 2026. Always confirm on the vendor’s site. Pricing page ↗
Open source
- Open source
- YesVendor claim[3]
Model weights are published on Hugging Face under Apache 2.0; downloading requires accepting an acceptable use policy. The technical report, training code (a Megatron-LM fork) and pretraining data reconstruction scripts are published on GitHub.
- Licence
- Apache-2.0 · OSI-approved licence
- Source code
- github.com/swiss-ai
- GitHub
- 137 public repositories
- last commit
Statistics from GitHub, fetched 15 Oct 2026.
Company
National research initiative launched in 2023 and co-led by EPFL and ETH Zurich, with the Swiss National Supercomputing Centre (CSCS), that develops open AI models such as Apertus.
Sources & verification
Every claim on this page is marked as verified, vendor claim or inferred. Unknown means we could not confirm it — not that the answer is “no”.
- Verified Confirmed by an independent or official source such as the commercial register.
- Vendor claim Stated by the vendor in its own documentation, website or legal terms.
- Inferred Derived by the editors from indirect evidence — see the note.
- [1]
- [2]
- [3]
- [4]
- [5]
- [6]
Swiss AI Initiative
- [1]
- [2]
- [3]