There are three ways to use Apertus, the open Swiss language model developed by the Swiss Federal Institute of Technology in Lausanne (EPFL), the Swiss Federal Institute of Technology in Zurich (ETH Zurich) and the Swiss National Supercomputing Centre (CSCS): an online chat service, a platform hosted in Switzerland, or downloading the weights from Hugging Face. The Apache 2.0 licence allows commercial use.

The current version, Apertus 1.5, was released on 24 July 2026, according to the CSCS announcement. It comes in two sizes, 8 and 70 billion parameters, according to its Hugging Face model cards and the Public AI blog post. It succeeds the first version of 2 September 2025. Which access route to choose mainly depends on one question: where should the data be processed?

Where can you try Apertus without installing anything?

The most direct option is the Public AI service. In its post of 24 July 2026, this organisation describes itself as the official international distributor of Apertus. Apertus 1.5 is available there on chat.publicai.co, on its developer platform platform.publicai.co and via Hugging Face.

There are two levels of access, according to a blog post by the company Liip dated 3 September 2026. The publicai.co/chat page offers the 8 billion version without registration. A free account on chat.publicai.co unlocks the 70 billion version, with the new “thinking” mode. Liip says this service is hosted under Swiss jurisdiction.

Liip also offers Apertus 1.5 in TextMate, its tool for proofreading, rewriting in plain language and translating into more than 12 languages. It can be tested at liip-textmate.liipgpt.ch. Finally, the Hugging Face model card for Apertus 1.5 70B includes a playground served by Public AI as the inference provider.

These trials are fine for getting started. For confidential data, it is better to use an access route where requests are not sent to a third party.

Which version of Apertus should you download?

The models are published by the swiss-ai organisation on Hugging Face, the distribution platform for artificial intelligence (AI) models. Version 1.5 goes by the names swiss-ai/Apertus-v1.5-8B and swiss-ai/Apertus-v1.5-70B. The first version remains online, for example as swiss-ai/Apertus-70B-Instruct-2509.

VersionSizesAccepted inputsMaximum contextTraining
Apertus 1.0 (September 2025)8 and 70 billion parametersText65,536 tokens15,000 billion tokens
Apertus 1.5 (July 2026)8 and 70 billion parametersText, images, audio262,144 tokens19,000 billion (8B), 17,000 billion (70B)

The figures come from the two Hugging Face model cards. Apertus 1.5 extends the training of version 1.0 with 4,000 billion tokens for the 8B and 2,000 billion for the 70B. It only generates text: no images, no sound.

For individual use, EPFL considers the smaller model better suited, in its 2025 announcement. The CSCS also points to Apertus Mini, a suite of 16 compact models obtained through distillation and quantisation. They are aimed at hardware with limited resources.

The model card for the first version also links to compressed versions that can be used in local applications such as llama.cpp, Ollama or LM Studio.

What does the Apertus licence allow?

Apertus is released under the Apache 2.0 licence, which allows both research and commercial use, according to the CSCS. EPFL specifies that the licence also covers the documentation, the training code and the intermediate checkpoints. An SME can therefore integrate the model into a product it sells, provided it complies with the usage policy.

Version 1.5 adds an extra step. The repository is public, but you must accept conditions to download the files. Users confirm that they have read the Apache 2.0 licence and the Apertus 1.5 Acceptable Use Policy. They also share their email address and Hugging Face username with the authors. A Hugging Face account is therefore required.

Two practical points emerge from the model card:

  • no output filter is provided to remove personal data covered by erasure requests;
  • the team says it handles these requests by updating the filters in its data pipeline and publishing new weights at least twice a year; it recommends downloading them every six months.

Requests to erase personal data or protected content go to llm-privacy-requests@swiss-ai.org and llm-copyright-requests@swiss-ai.org. The documentation required by the EU AI Act is published on the same model card.

Which Swiss platforms host Apertus?

Swisscom, a strategic partner of the Swiss AI Initiative, has opened Apertus to its business customers since 2 September 2025, according to EPFL. On the page for its Swiss AI Platform offering, the operator lists Apertus among the models on its managed inference endpoints, accessible through its Digital Marketplace. It promises that customer data stays in Switzerland.

The academic world has its own access. The CSCS has launched an inference service for Apertus 1.5, intended for all its users and, by extension, the Swiss academic community. It provides a dedicated application programming interface (API) endpoint.

According to Liip, the provider Infomaniak also offers the 70 billion parameter version of Apertus 1.5 through its API. Liip mentions it for those who do not want to manage inference themselves.

The Swiss AI Initiative, for its part, says it wants to make its models usable by Swiss players, including SMEs and start-ups. Our AI tools and models section follows new offerings.

What hardware do you need to run Apertus yourself?

Installation is straightforward for a professional, according to EPFL, but real-world use requires servers, cloud infrastructure or a dedicated interface. The model on its own does not provide a chat window.

The Apertus 1.5 model card gives the commands for the vLLM serving software. The one for the 70B splits the model across four graphics processing units (GPUs). The one for the 8B does not split it. The model card advises reducing the context length if memory runs short.

As of 21 November 2026, support for Apertus 1.5 has not yet been integrated into the official releases of vLLM and Transformers, according to the model card. The team provides a modified Docker image for the amd64 and arm64 architectures. The first version, for its part, works with Transformers 4.56.0 or later.

For a personal computer, Liip mentions the 8B: it runs comfortably on recent hardware, with the mlx-lm tool on an Apple silicon Mac. Compressed, it has even run on an 8 GB laptop graphics card, according to users quoted by Liip.

Other models target the same modest hardware. The French cloud provider OVH claims that its Olala Mini model, with 7 billion parameters, runs on an 8 GB GPU, according to Inside IT on 20 November 2026. The Lausanne start-up Giotto, meanwhile, designed its Solo model for a single GPU, as noted in our article on Giotto.

What are its limits compared with the big commercial models?

The authors do not claim parity. “Apertus does not seek to compete with the frontier models of private companies,” says Imanol Schlag, a researcher at ETH Zurich, in the CSCS announcement. The goal is a transparent foundation on which to build.

The model cards point out that Apertus may produce inaccurate, inconsistent or biased content. It should serve as an assistance tool, not as a definitive source. The “thinking” mode is disabled by default and must be enabled manually.

Some inputs are only partially supported. Video is not supported natively: frames have to be extracted from it. Voice processing remains experimental, according to Public AI.

On dialect, the Intochat test published in October 2026, summarised in five AIs tested in Swiss German, only covered models from Anthropic, Google and OpenAI. It therefore says nothing about Apertus.

Key figures

  • 2 sizes: 8 and 70 billion parameters, for both Apertus 1.0 and Apertus 1.5.
  • 262,144 tokens of context for Apertus 1.5, four times more than version 1.0.
  • 16 compact models in the Apertus Mini suite.
  • 4 GPUs in the vLLM command proposed for the 70B.
  • 8 GB: the laptop graphics card on which the compressed 8B ran, according to Liip.
  • 6 months: the recommended interval for downloading new weights.

What we don't know yet

  • Swisscom does not publish a price for Apertus on its Swiss AI Platform page, which refers visitors to a sales contact.
  • Public AI announces Apertus 2 “in months, not years”, without a specific date.

Sources