European Public Services Test Open AI Models Across Healthcare, Administration and Local Languages

European public service teams are testing open models in settings where control over data, infrastructure and local languages shapes whether an AI system can be used at all. The projects span medical imaging, employment services, document search, collaboration software and voice interfaces. Their shared engineering problem is practical: build useful applications while keeping deployment and evaluation close to the organizations that use them.

Open models give teams more room to inspect, adapt and run systems on infrastructure they control. Some also publish information about training data, which can help organizations assess licensing, bias and other data concerns. That flexibility still leaves work to do: teams must evaluate model quality, set safeguards, manage operating costs and determine whether systems perform consistently for different communities.

Several projects are pairing these models with localized synthetic personas. NVIDIA’s Nemotron-Personas collection supplies partner-grounded datasets for NeMo User Sim, an open source library for simulating users at population scale. The aim is to evaluate and train models and agents in local languages using simulated users, rather than relying on private records of real people.

Shared tools for building and governing applications

Berlin-based deepset is expanding its Haystack Enterprise Platform, which underpins AI@EC, the European Commission’s shared platform for developing and running AI applications on infrastructure managed by the commission.

The expanded platform connects agent development and execution with optimization and governance. It adds ways to compare operating costs, assess answer quality and trace problems. NVIDIA Triton Inference Server supports shared services that retrieve and rank relevant information. AI@EC deployments use reusable components, role-based access, audit trails and safeguards for applications such as document summarization, knowledge management and analysis across languages.

The design points to a wider operational lesson: a model alone does not make a service ready for use. Reusable infrastructure and access controls help teams manage applications, while quality checks and problem tracing give them ways to inspect how those applications behave. The source does not provide performance results or cost figures, so the practical gains remain to be measured in each deployment.

Healthcare and employment services in Poland

Deloitte is helping Poland develop agentic systems for healthcare and employment services using open models. The eHealth Centre is building an Intelligent Services Platform for medical imaging with NVIDIA DGX BasePOD and the MONAI framework. The planned tool will help physicians analyze X-rays, CT scans and MRI images, identify abnormalities and prioritize urgent cases, with the goal of shortening waits for results.

Separately, Deloitte is supporting an agentic factory for more than 400 labor offices. It will run on NVIDIA accelerated computing infrastructure and use models including PLLuM, GPT OSS and Inkling for document search, summarization and digital assistants. The intended role is to help staff assist job seekers more efficiently. No accuracy, throughput or service-time results are reported, leaving those measures as key evidence to watch as the systems develop.

NVIDIA is also working with Poland’s SpeakLeash and the University of Ljubljana in Slovenia on country-specific additions to its synthetic personas program. Localized datasets are planned to follow.

Local data and interfaces across Germany and the UK

London-based Locai Labs has released Nemotron-Personas-UK, a dataset of 1 million synthetic personas designed to reflect demographic and geographic diversity across the UK. Proposed uses include evaluating healthcare systems and checking for differences in model behavior across age groups, education levels, and rural and urban populations. For developers, the dataset offers a structured way to probe variation across simulated communities; it does not by itself establish how a model performs with real users.

Germany’s Materna is introducing Sinfonie, an on-premises platform combining generative AI, voice interaction and 3D avatars. It can use NVIDIA Audio2Face for animation. Built on a project for a German state administration, the platform lets employees ask spoken questions to find institutional knowledge while keeping data on their organization’s infrastructure.

For work involving classified information, secunet, NVIDIA, Red Hat and deepset are developing a cloud offering based on secunet’s SINA Cloud foundation. The planned stack combines NVIDIA accelerated computing and Nemotron open models, Red Hat OpenShift AI for managing AI workloads, deepset’s Haystack agent orchestration and infrastructure operated by secunet in German data centers. The source says the foundation includes security components authorized for classified use; it does not describe the offering’s availability or performance.

The Sovereign Open Source Foundation Models research initiative, or Soofi, is developing models for German industry and public services. Its work includes Nemotron-Personas-Germany: an open dataset of 1 million records, each containing six synthetic personas. The personas are grounded in the 2022 census and cover all 16 federal states and 400 districts. Created with NVIDIA NeMo Data Designer without private data from real people, the dataset is intended to help teams check how systems behave across regions, age groups and communities. It can also be used with NeMo User Sim to evaluate and train models against simulated German users.

Collaboration tools in France and Greek-language systems

France’s public purchasing body for digital services, CANUT, selected LINAGORA and Docaposte to provide a collaborative workspace based on Twake.ai. More than 2,500 eligible organizations—including municipalities, schools and healthcare institutions—can use the framework without arranging their own tender. The workspace includes email, document sharing and collaboration tools, with NVIDIA technology supporting speech transcription and hosting in France. Région Occitanie is the preferred partner for scaling deployment.

Greek AI and energy company KIEFER is developing an agentic factory on NVIDIA DGX SuperPOD with DGX GB200. It combines Sophea open-weight models with Greek-language tools for information retrieval, speech and digital assistants. The focus is access to institutional information for Greek-speaking teams, with organizations controlling where agents run and how data is handled.

Compute and the next application layer

NVIDIA Open Models Codefest brings startups, software providers and research labs together to build applications for public service needs, including medical reasoning, AI tutoring and model governance. The five named teams are Pleias, Zehnlab.ai, Disseqt, Algoritmica GmbH and ABA Technology. Deutsche Telekom and T-Systems supplied compute infrastructure for training and deployment.

Across these efforts, the strongest common thread is local control paired with evaluation: teams are assembling compute, models, datasets and application tools around specific languages and service needs. Synthetic personas may make it easier to test demographic coverage, and shared platforms can give teams repeatable ways to build and inspect applications. The next useful evidence will be measured performance, operating costs and outcomes in real workflows.

Read all NVIDIA in Europe announcements. Learn about NVIDIA’s work on secure AI and agent development. See the notice regarding software product information.