Apertus AI: the open Swiss LLM, explained for SMEs

Apertus is a large language model, or LLM: a program that reads and writes text. It is developed by the Swiss AI Initiative, which brings together EPFL, ETH Zurich and CSCS, the Swiss National Supercomputing Centre. Its name means “open” in Latin.

The first version came out in September 2025, in two sizes: 8 and 70 billion parameters. Parameters are the model’s internal settings; the more it has, the more computing power it needs to run.

Apertus 1.0 was trained on 15 trillion tokens, the small pieces of words a model reads, using 4,096 GH200 graphics processors. About 40% of this data is in languages other than English.

Apertus 1.5 keeps the same two sizes, 8 and 70 billion parameters. Its training was extended by 4 trillion tokens for the 8B and by 2 trillion for the 70B.

• A context of 262,144 tokens, four times that of 1.0: the model reads much more text in a single request, for example several long documents.
• A “thinking” mode, switched on when needed, in which the model reasons before it answers.
• Image understanding, and audio input that is still experimental.
• Better instruction following and better tool use, meaning the ability to call a function or an API in the middle of a task.

The Apertus weights are published. They are the values that make up the model, and with them you can run it on your own machine. The recipes for the training data and the details of the training are public too, so anyone can check how the model was built.

The Apache 2.0 licence allows commercial use. An SME can build Apertus into an internal tool or into a service it sells to its customers.

To assemble the training data, the team respected the opt-outs that websites expressed in machine-readable form, and removed personal data. No output filter for personal data is supplied, so that check has to be built into the tool that uses the model.

Apertus can be tried in a browser first, then installed wherever your data has to stay.

• Public AI (chat.publicai.co): a chat window, with the 8B available without an account and the 70B with a free account.
• Hugging Face: download the swiss-ai/Apertus-v1.5-8B and swiss-ai/Apertus-v1.5-70B models.
• Locally: Ollama, mlx-lm or vLLM, tools that run a model on a computer or a server.
• Via API: Infomaniak offers the 70B, which you can call from your own software.

Apertus makes sense when the model has to be verifiable from end to end, when it has to run on a machine you control, or when your texts move between Switzerland’s national languages. Data control is covered in detail in the article on sovereign AI assistants and models.

Lumineth uses Claude, from Anthropic, for analysis and writing. The model is chosen task by task: every candidate is given the same examples taken from your business, and the one that gets the answers right is kept.

A well-built tool stays reversible. The instructions, the checks written in code and the reference cases stay in place, so the model can be replaced and then tested on the same cases. The full framework is on the AI agents in Geneva page.