User Goals
- Scale our embedding model into a full cross-modal foundation model for construction data.
- Prove that domain-trained, EU-models can outperform large general-purpose ones.
- Give the European Architecture, Engineering, and Construction (AEC) sector a sovereign semantic infrastructure layer.
Current State and Key Challenges
flinq is a Stuttgart-based deep tech startup and is currently a technology spin-off (TGU) of TTI GmbH at the University of Stuttgart. We are training a domain-specific, cross-modal embedding model for the European construction industry — an industry that generates enormous amounts of data every year, including tenders and an increasing number of Building Information Modeling (BIM) models and product catalogs. As a cross-modal model, our embedding model processes many data types simultaneously, converting BIM models (IFC), tenders (GAEB), manufacturer data, construction site images, and other construction-specific formats into a common vector space. We make this available via an OpenAI-compatible API.
Our finely tuned 600-million-parameter Proof of Concept (PoC) model shows promising benchmark results compared to OpenAI’s “text-embedding-3-large” and the “Qwen3” model, which is 13 times larger. Initial tests also show that 3D geometry and text can be embedded in the same space. This approach could address the challenge of integrating diverse construction data.
To date, flinq has run all training on a single rented GPU. For scalability and robustness, we require more computing power than we currently have, highlighting the need for access to HammerHAI’s computing infrastructure.
Support from HammerHAI
We are looking for continuous access to GPUs that can scale with us as we transition from today’s fine-tuning experiments to the significantly larger training runs for new models that lie ahead. In addition to raw computing power, HammerHAI’s expertise in efficient large-scale training and its connections to the broader European AI ecosystem are also of great value.
Intended Value and Impact
Using larger, fully cross-modal models, we aim to make flinq the semantic interoperability layer that the construction industry lacks, connecting BIM, tenders, products, and cost data. While the field moves toward horizontal, general-purpose models, we want to establish flinq as a dedicated embedding model for the construction industry and a trusted, vertical, EU-hosted solution. This will provide a common foundation that will reduce the need for time-consuming manual data reconciliation.
Current Status within HammerHAI and Next Steps
Our collaboration with HammerHAI is still in its early stages: Onboarding is complete, and we have access to KISSKI (AI service center for critical and sensitive infrastructures). We will soon begin our first training runs, including initial scaling experiments to prepare for our next milestone: successfully running a four-billion-parameter (4B) model. At the same time, we are expanding our data collection and continuing development of our data strategy for the larger, cross-modal dataset.
Contact
Matthias Hornung, flinq (TTI GmbH, Universität Stuttgart), info@flinq.ai
