The Suniworks journal

Apertus: what Swiss SMEs can use the open AI model for

Apertus gives businesses an open foundation for writing assistants, translation and document tools. Understand what you get, how to access it and what to check before using it with company information.

Updated on .

A workstation connected to a server processes business documents, while a separate document awaits human approval.

Apertus can provide the engine for an assistant that drafts messages, translates text or helps explore documents. Developed by EPFL, ETH Zurich and CSCS, it gives Swiss SMEs a model they can inspect and adapt, with access through applications or a custom-built tool. Start by assessing a specific task rather than assuming it will replace your existing software.

The essentials before you try it

  • Apertus supplies the model behind an application. Your documents, user accounts and business processes need a separate application layer.
  • Its published components and Apache 2.0 licence support inspection, adaptation and commercial use.
  • A useful trial needs a defined task, an expected answer and someone responsible for checking the result.

What does a business actually get with Apertus?

A language model is a system trained to process and generate text in response to a request. Apertus supplies that foundation. An application built around it provides the practical features employees interact with, such as a conversation window or a way to submit documents.

This distinction matters when choosing how to try it. Downloading the model does not give your team a finished workplace assistant. In its launch announcement, EPFL explains that practical use also requires computing infrastructure and an appropriate interface. For a business without developers, an application that handles those elements is the more practical starting point.

What does its openness let you do?

The project makes the model's learned parameters, training code, training data and development methods accessible. Parameters are internal values adjusted during training; they form an essential part of the model files. This gives a technical team material it can examine when assessing how Apertus was built or adapting it for a particular purpose.

A licence sets the conditions under which something can be used. Apertus is distributed under Apache 2.0, which permits commercial applications as well as research. A business can therefore build a product or internal tool around it, subject to the licence conditions. The practical benefit is access to the underlying technology for examination and adaptation.

Which versions and languages matter for your project?

The initial release offered models with 8 billion and 70 billion parameters. EPFL described the smaller model as more suitable for individual use. These sizes identify technical variants; they should not be treated as a quality rating for accounting, administration or customer service.

For Apertus 1.5, EPFL announces image and audio understanding alongside text, plus improvements in reasoning, instruction following and the use of external tools. Tool use means an application can connect the model to other functions, provided those connections have been developed. The Apertus 1.5 announcement also describes Apertus Mini, a family of sixteen compact models intended for environments with fewer computing resources.

Ask which variant an application actually runs and which capabilities its interface exposes. A model's ability to process audio, for example, does not establish that every application using it accepts audio files.

How useful is its multilingual training in Switzerland?

According to EPFL, the initial training material covered more than a thousand languages, with approximately 40% of the material coming from non-English sources. Swiss German and Romansh were among the languages represented. That makes Apertus worth assessing for a company working across Switzerland's language regions.

For a French-speaking Swiss SME, the useful test is a request in the languages and terminology its staff actually use. Include professional expressions and exceptions that matter to the business. Have a competent speaker review translations before they become quotations, contractual communications or customer commitments.

What could you ask it to do in an SME?

The following are hypothetical trials, not customer results. They apply the writing, translation and document-related uses described by EPFL to everyday business situations.

An accounting firm: draft a request for supporting documents

An accounting firm could give an assistant a staff-approved list of required documents and ask for a clear client message. Use a fictional case first. Check whether the draft preserves every required item and introduces any additional requirement. This evaluates writing support without asking Apertus to determine the tax treatment of a client's affairs.

A professional practice: simplify an administrative procedure

A practice could ask an assistant to turn a non-confidential procedure into a reference sheet for reception staff. The person responsible would compare the draft with the original, checking conditions, exceptions and responsibilities. For a medical practice, keep this initial trial within administration and leave clinical decisions outside its scope.

A retailer: prepare a reply in another language

A retailer could use its public returns policy to draft a response to a hypothetical customer. Check that the translated reply preserves the shop's conditions and adds no promise. This creates a focused language task without starting with order histories or customer contact details.

Can you use it without technical skills?

Yes, through an interface that incorporates Apertus and handles its operation. EPFL's initial announcement describes access through deployment partners and identifies Public AI as an access route. In that arrangement, employees use a service built around the model rather than working with the model files themselves.

Before selecting an interface, ask which model it uses, what features are available and what conditions apply to submitted documents. For the separate decision about remote services, Swiss hosting or an internal installation, use our guide to where company data is processed.

What would running it yourself involve?

EPFL makes the models available through Hugging Face, a platform for distributing models. Running Apertus then requires computing resources, an interface and the technical skills to operate the resulting system. Ask your technical lead to assess a named variant against the intended documents and user workload before buying equipment.

The model download is free. Keep that separate from the budget for infrastructure, integration and ongoing operation. Access to the model files gives you a foundation to build on; an operational application still needs those additional elements.

How should you judge it against major commercial models?

The Apertus team presents its project as a transparent foundation for organisations, rather than an attempt to outperform the most advanced private models. EPFL reports improvements in Apertus 1.5, but those announcements should not be read as a promise that it will produce the best answer to your business task.

For an SME, a useful comparison is a controlled trial using identical requests and expected answers. Record incorrect statements, missing information and the changes a reviewer must make. Include the checking effort in your assessment: readable wording alone is not enough to approve a message that commits the company.

Keep approval with a responsible person during the trial. Before introducing another workplace assistant, also review the broader issue of control over company information, so that staff have a clear rule about what they may submit.

Where to start this week

  • Select a narrow drafting or translation task with an answer your team can check.
  • Prepare public or fictional examples, the expected content and the errors that would make a result unacceptable.
  • Identify an interface offering Apertus, then review its model choice and document-handling conditions.
  • Compare its output and review effort with your current method before deciding whether to continue.

If the trial supports further work, define the data boundaries before commissioning an integration: which documents may enter the application, who may access them and which actions need human approval. Make those decisions part of a documented approach to keeping control of your business data.

A software decision to prepare? Tell us about the context and the work your application needs to support.

Discuss your project