April 23, 2026 -
Which technology trends are shaping the future of large language models (LLMs) – and what makes Europe’s role unique? In this interview, Jan Plogsties, Head of the Generative AI Department at Fraunhofer IIS, explains what the institute is currently researching and where those developments are heading.
Jan Plogsties: One key question is how to run AI models locally on small devices – on smartphones, in vehicles, or in industrial systems. More specialized accelerators and processors are entering the market, particularly for smartphones and embedded systems. This makes it possible to run LLMs locally – without the cloud and without latency.
However, these models then must operate with significantly fewer parameters and far less computing power. At Fraunhofer, we refer to as this “large models for small devices.” This is particularly exciting for interactive applications: for example, if I give a spoken command to a machine, the response must come almost in real time and cannot first be routed through a large cloud-based model. We are addressing this challenge, among other things, in the Bavarian project DSgenAI. With our model ELMOD, we have already demonstrated that a high-quality German language model can run locally on a smartphone.
Another exciting project we are currently working on is SOOFI. In this project, we are developing a family of language models with sizes of up to 100 billion parameters, together with some of Germany’s leading experts in LLM training. To ensure that the model ultimately delivers what is needed in real‑world applications, we involve industry partners who contribute their requirements directly. With SOOFI, we aim to help establish sovereign AI models in Germany and Europe.
Jan Plogsties: Because we truly need to understand and master this technology ourselves. In a world where international dependencies can become tangible at any time, we cannot rely on the permanent availability of core services from other regions. Technological independence is therefore a strategic necessity.
In addition, Europe has a strong AI research landscape, but so far, we too rarely succeed in turning this knowledge into powerful commercial successes. To retain talent and expertise in Europe over the long term, we need to build and continuously develop our own technologies.
Beyond that, developing our own models allows us to shape AI according to European ideas and values. We can deliberately define which languages and societal perspectives should be taken into account. Data filtering, bias control, and regulatory guardrails are integrated during development from the very beginning. This creates a high level of trustworthiness and cultural diversity. Our first LLM, Teuken‑7B, for example, deliberately includes 21 European languages and is not only available as open source, but also complies with EU standards regarding the AI Act.
Jan Plogsties:There are many companies that already use LLMs to make everyday life easier for people and to improve efficiency in office work. For such use cases, there are numerous offerings on the market – from large corporations as well as open‑source models. That is useful, but the real potential of LLMs lies in completely different areas: for example, in the development of new active ingredients for the pharmaceutical industry or in technical simulations in the automotive sector. Conventional language models also reach their limits in use cases where sensitive or highly specialized data needs to be processed – for instance, when dealing with tax data, legal questions, or government tasks.
Because we also develop models ourselves, we possess deep expertise in the specialization and optimization of language models. We decide which training data flows into the language model and maintain control over data quality. Base models can be specifically enriched with industry- or company-specific knowledge. In this way, we develop language models for our partners that are precisely tailored to their industry, infrastructure, and compliance requirements.
Thank you very much for the interview.