Swiss AI model Apertus gains users but still trails leading rivals

More than four million downloads have put Switzerland’s Apertus firmly on the global open AI map, one year after its launch. The figure comes from Hugging Face, the platform where developers and researchers share machine learning models. The project team says it has also identified at least 70 external deployments and more than 100 derivative versions.

Those numbers show reach, though they do not prove that millions of people use Apertus regularly. Downloads can represent experiments, evaluations or abandoned projects. The more telling evidence comes from organisations that have adapted the model for specific needs, including Singapore’s SEA-LION project, which combines Apertus with other systems for Southeast Asian languages.

Apertus emerged from Swiss research institutions with an ambition to offer a fully open, multilingual model shaped by European data protection principles and the European AI Act. That positioning gives Switzerland a visible role in a field dominated by US and Chinese companies.

The model has gained users and technical upgrades. It still trails leading international systems on important tasks, limiting its appeal where reliability, reasoning and broad general performance matter most. The result is a model with growing adoption, but an unfinished competitive case.

Apertus 1.5 arrived in July 2026 with image understanding and improved reasoning capabilities, giving the Swiss project a stronger technical base than the version released a year earlier. The official roadmap points to Apertus 2.0 in 2027.

The release reflects the pace of the AI market, where models are updated continuously and users expect rapid gains. Apertus has had to address early criticism that it struggled with basic questions. Some users described the first version as “unusable”, according to the source article, a reception that exposed the distance between a high-profile launch and dependable everyday performance.

Oleg Lavrovsky, the model’s community lead, argues that comparisons with ChatGPT miss the project’s purpose. Apertus was created as a foundation model that businesses, public institutions and researchers can inspect, modify and deploy for their own requirements. Its weights, training code and information about the training data are intended to be publicly available, distinguishing it from open-weight systems such as Meta’s Llama and Alibaba’s Qwen.

That design offers control and transparency. It also shifts responsibility to adopters, which must improve, test and maintain their own applications. For organisations seeking a ready-made assistant, that extra work can make Apertus less attractive than a polished commercial rival.

Multilingual performance is emerging as Apertus’s clearest practical advantage. Ticino-based Artificialy says it uses the model extensively in work for clients, particularly for translation and multilingual conversations. Damiano Binaghi, the company’s head of deep learning, says the decision is commercial rather than symbolic: “We’re not doing charity. We don’t use Apertus because we have to, but because it’s a good model.”

That assessment matters in Switzerland, where public services and businesses operate across German, French, Italian and Romansh, and where cross-border communication is routine. A model that can be adapted and run under a client’s own control may appeal to Swiss companies and public bodies handling sensitive information.

Apertus’s open structure also permits local adaptation. SEA-LION’s work in Singapore demonstrates how a model developed in Switzerland can become part of a broader language technology stack. The project independently adapted Apertus and other models for multiple Southeast Asian languages.

Still, a strong result in translation does not automatically solve wider weaknesses. Organisations must assess accuracy, security, cost and maintenance for each use case. The available evidence points to targeted value rather than universal superiority. Apertus is gaining ground where multilingual adaptation and transparency carry weight, while leading international systems retain an advantage across many general tasks.

Apertus now has evidence of adoption, but not yet proof that it can match the leading systems. The four-million-download milestone and the reported deployments establish interest. They do not provide a benchmark for reliability, nor do they show how many organisations have moved from testing to sustained production use.

Lavrovsky acknowledges the uncertainty over how much of Apertus’s potential is translating into tangible value. That gap will matter as Swiss companies, universities and public authorities decide whether openness justifies the technical work required to deploy the model. Open access can support auditability, customisation and data control, but those benefits come with demands for in-house expertise.

Switzerland’s next test will be practical. The project must show that future releases improve performance on the tasks that exposed weaknesses in the original model, while preserving the multilingual and open characteristics that distinguish it. The 2027 Apertus 2.0 roadmap gives developers a clear point at which to assess progress.

For now, Apertus occupies a credible but limited position. It is a Swiss research platform with international users, real deployments and a growing ecosystem of adaptations. Its future influence will depend less on download totals than on whether those users continue to build on it and whether the resulting applications deliver dependable results.