Thursday, April 30 2026

Creating the conditions for enterprises to truly own and trust the models they put into production: A conversation with Dr. Nicolas Flores-Herr

12:37

Deploying a model in a demo is straightforward. Putting one into a production workflow that your organization genuinely relies on – in healthcare, public administration, industrial operations, or financial services – is a different problem entirely. It requires models you can audit, training data you can document, and evaluations that map to the regulatory and operational requirements of your specific context. It requires, in short, AI you can trust not just on benchmarks but in practice.

That level of trust doesn't come from a vendor's assurance – it comes from verifiable artifacts: documented training data, transparent model cards, and evaluation suites you can interrogate against your own compliance requirements and business use cases. The organizations pulling ahead are the ones building the internal capability to systematically track, compare, and evaluate models against their specific needs – from operational performance to regulatory fit – so that putting a model into production is a deliberate, auditable decision rather than an act of faith in a vendor's roadmap.

One of the most noteworthy approaches to closing the gap between SOTA research and enterprise-ready AI is happening in Germany, where a new generation of consortia is structurally connecting lab-grade model development with the real adoption constraints of industry. To learn more about how this works in practice, I sat down with Dr. Nicolas Flores-Herr – who has spent the last several years leading the development of open source foundation models at Fraunhofer IAIS, coordinating across research institutions, startups, and industry partners including NVIDIA and Deutsche Telekom to bring production-grade AI from the lab to enterprises of all sizes in Germany and beyond.

My conversation with Dr. Nicolas Flores-Herr

Gaël Varoquaux: Nicolas, you've spent the last few years leading one of Germany's most ambitious efforts to build open source foundation models. In Soofi alone, you've brought together 7 research institutions, 2 startups, industry giants like NVIDIA and Deutsche Telekom, and a federal ministry under one roof to work together to build an open source foundation model. Before we dive into your work, tell me, where does your drive come from to make sure that organizations – from the federal government of Germany all the way to German Mittelstand enterprises – can genuinely own the AI tools they use and depend on?

Nicolas Flores-Herr: My drive comes from a simple observation. The organizations that built Germany's economic strength over decades, from federal agencies to mid-sized industrial leaders, now depend on AI systems they neither own nor can audit. In a stable geopolitical environment that might be acceptable. Today, it is an immediate strategic risk. The companies and institutions that learn to "surf the wave" will turn AI into productivity gains measured not in percentages, but in multiples. Given that scale of impact, this is not an ideological preference for open source over closed systems. It is the question of whether European institutions and enterprises retain the technical agency to develop, govern, and continue the AI capabilities they rely on. That, for me, is what the current conversation about “sovereign AI” is really about.

Gaël Varoquaux: I know you focus on open source foundation models but first I want to zoom out for a moment, because I think there's a broader point about what it means for an organization to truly own and trust its data science and AI. In your experience working with industry partners, how much do enterprises understand that owning their AI isn't just about which foundation model they run, it's also about the full stack they build their workflows on?

Nicolas Flores-Herr: You are right that the discussion has matured. A year ago, enterprises asked which model to license. The more advanced ones now ask: who controls the data pipelines that adapt it to our domain? Who runs the post-training? Where does the compute sit? The center of gravity is shifting, and agentic AI is accelerating that shift. Workflow logic that used to be hard-coded into applications – like when to call a tool, how to sequence a task, when to escalate – and is increasingly handled by the model itself, while applications become the layer that provides the tools and data the agent reasons over. That is why owning your AI means owning the layers where capability is actually created: the models and the tools and data that go into training, finetuning, and evaluating them before putting them into production. The full stack follows from that, not the other way around.

Gaël Varoquaux: Your team builds both open source foundation models and tailored applications for enterprises. Can you talk me through how you're facilitating synergies between the supply of your R&D in the lab, and demand for these tools in industry?

Nicolas Flores-Herr: In Soofi, the answer is structural rather than promotional. We built the consortium so that pre-training, post-training and application sit in the same project: nine partners spanning Fraunhofer institutes, DFKI, Germany’s top AI universities, startups, and industry. That removes the classical research-to-industry handover, which is where most European R&D loses its momentum. European applied research is uniquely positioned to operate this loop, and that is where investment should go.

Gaël Varoquaux: I'd like to pick your brain a bit more about enterprise adoption and impact of the tools you and your team have built. Where are you seeing traction with open source foundation models in German enterprises right now – which sectors, which organizations, which use cases, and why?

Nicolas Flores-Herr: Traction is encouraging and broad. We see real momentum in regulated sectors, public administration, healthcare and financial services, where auditability and data governance are non-negotiable. We see equally strong demand from German SMEs, from machinery and chemicals to automotive and energy, where domain expertise is the moat and adapting models to it requires deep collaboration. A third cluster is emerging around agentic AI and process automation across both. The common pattern is clear: wherever research and industry need to co-develop complex AI solutions, scaling and transfer become the bottleneck. That is where European applied research has to step up, we also need to re-think national and European funding programmes.

Gaël Varoquaux: And what are the key blockers for enterprise adoption? What's the thing that, if you could fix it tomorrow, would unlock the most adoption by enterprises?

Nicolas Flores-Herr: If I had to pick one, it would be the gap between how rapidly model capability is advancing and how slowly enterprise software architectures are adapting. Many organizations still treat foundation models as one component in a traditional stack, surrounded by rigid platforms and brittle integrations. That view is becoming obsolete. The same sentence appears above already. Platforms, applications, and workflows must reorganize around what the AI can do. The good news is that this is a solvable problem. Enterprises that internalize this shift this year will be the early winners.

Gaël Varoquaux: I'm curious about your current thinking about open source as a vector for competitiveness in AI. There's a recurring debate about whether open-sourcing AI tools and models gives away competitive advantage or actually accelerates it. Based on your experience with projects like Soofi, TrustLLM, and OpenEuroLLM, what's your view on that tradeoff?

Nicolas Flores-Herr: My honest view, after Soofi, TrustLLM and OpenEuroLLM, is that the framing has moved on. While I believe that Open Source is Europe’s only chance to build value-creating AI ecosystems, we should move the debate from open source versus closed to new topics. First, continuity of model building: who guarantees a credible V2 and V3, not just a single release from a one-off research project? Second, ownership: who actually holds the equity in the supposedly European provider behind a model? In today's geopolitics, those questions will dominate procurement decisions. What Europe needs is an open AI ecosystem in the sense that no critical technical dependency runs through a single external party, serving industry, the public sector, and Public AI alike. That is the real competitive vector.

Gaël Varoquaux: Foundation models and agents are increasingly being deployed in enterprise workflows. The productivity gains are real, but so are the concerns about trusting outcomes. In your opinion, what can developers of open source software and open source foundation models like you and I do to increase trust when these tools are used in enterprise settings? And are there particular sectors where getting that trust right feels most urgent to you?

Nicolas Flores-Herr: Trust comes from verifiable artifacts, not claims, and there are really two layers of it. The first is sector-specific compliance: in healthcare, public administration or financial services, you need documented training data, clear model cards and evaluation suites that map to regulatory requirements. Open foundation model developers can deliver exactly that. The second layer is universal and underappreciated: provenance. Can you trust the organization that built the model? With closed-source agentic systems, the concern is no longer only data leakage. It is the leakage of process knowledge, of how your business actually runs. That risk applies in every sector, not just the regulated ones.

Gaël Varoquaux: If you had to give one piece of advice to a CTO at a major European enterprise who is deciding right now whether to build on proprietary models or open source alternatives – what would you tell them, based on what's gone well and what hasn't?

Nicolas Flores-Herr: Don't frame it as proprietary versus open source. Frame it as: where in your stack can you afford a critical technical dependency on a single external provider, and where can you not? For commodity assistants, proprietary models are often the pragmatic choice. For the workflows that define your competitive position, you need a model and a post-training stack you can adapt, audit, and continue running independent of any single vendor's roadmap. That usually points to open foundation models. The mistake I keep seeing is treating this as a price-performance question. In my opinion, it is a question of sovereignty and organizational resilience, and it should be recognized as such by the leaders who want to be in the driver’s seat of their organization’s future.

About Dr. Nicolas Flores-Herr

Nicolas is the Manager of Foundation Models & Generative AI Systems and the Site Lead of Fraunhofer IAIS in Dresden, Germany, where he heads an interdisciplinary team working since 2022 on the development and application of sovereign and open source large language models (LLMs) and generative AI. Their expertise spans the full lifecycle: from dataset creation, pretraining and continued pretraining to fine-tuning and instruction tuning. They carry out this work in European flagship projects such as Soofi, TrustLLM, OpenEuroLLM.

In his leadership role, Nicolas combines scientific excellence with strategic impact. His team develops bespoke GenAI solutions for industry partners across diverse sectors, including scalable enterprise chatbots and specialised tools for semantic search and knowledge access.

With a PhD in physics and ten years of research in neuroscience and cellular physiology, Nicolas brings an analytical perspective on complex systems. Today, he applies this background to transform AI innovation into strategic business value, with a focus on scalability, digital sovereignty and sustainable impact.