Hafnia bridges the gap between data, training, and deployment, allowing developers to reduce dataset bias while training best-in-class vision AI models and deploying them in Smart City solutions. The result: AI systems that move beyond reactive learning and toward proactive readiness across their entire lifecycle.

Nvidia Physical AI Data Factory Blueprint is the reference architecture that unifies data curation, augmentation, evaluation, and agentic orchestration at scale. Powered by Nvidia Cosmos open world foundation models and Nvidia Osmo, it transforms raw real-world and synthetic data into high-fidelity, physics-aware, model-ready training datasets—accelerating development with speed, scale, and reliability.

Training beyond historical data

“Together with Nvidia, we are taking Hafnia to the next level by combining trusted real-world data with synthetic augmentation,” said Edward Mauser, Director of Hafnia at Milestone Systems. “This enables developers to train AI models that are not only accurate in known situations, but also resilient in the unexpected.”

AI systems typically learn from past events. But real-world environments, like cities, are unpredictable. Rare weather conditions, unusual traffic patterns, or region-specific vehicle types are often underrepresented in traditional datasets.

Integrating synthetic data

Hafnia addresses this gap by integrating synthetic data into its curated, real-world video library. Synthetic augmentation through Nvidia Cosmos Transfer allows developers to have access to data including rare or dangerous situations, balanced underrepresented object classes, model regional and environmental variations, and systematically reduced datasets biases.

Importantly, synthetic data doesn’t replace but enhances Hafnia’s real-world foundation. This ensures authenticity, compliance, and consistent annotation quality while expanding scenario coverage.

At the GTC event, Milestone will preview its upcoming Training-as-a-Service offering. Instead of piecing together fragmented pipelines, developers now have a seamless bridge from Hafnia's data to robust training infrastructure, allowing them to focus entirely on building high-performing video analytics.

TaaS gives developers streamlined access to the compliant, high-quality video data within Hafnia’s library – both real-world and synthetic data sets - that can be used within their own training pipelines. They will be able to customise datasets and fine-tune models for specific use cases. Because the data within Hafnia’s library is compliantly sourced and fully traceable, developers can train models with the confidence their model training is compliant with relevant regulations.

By removing the complexity of sourcing and managing training data, Hafnia allows developers to focus on building high-performing analytics solutions up to 30 times faster.

Visual language models

In partnership with Nvidia, Milestones Hafnia also offers VLM-as-a-Service: a suite of Visual Language Models built on Nvidia Cosmos Reason models and optimised for Smart City environments.

At GTC, Milestone announced the availability of a new EU-optimised VLM for traffic, already running with selected EU cities as customers. Additional models are coming soon, expanding the suite of VLMs offered through the VLM-as-a-Service platform to cover more smart city scenarios.

These hosted models break down the barriers to powering Computer Vision products with Generative AI solutions tailored for Smart Cities, eliminating costs for data collection, repeated retraining, and infrastructure scaling.

For teams looking to accelerate the final stages of the lifecycle by 70 times, Hafnia’s VLM-as-a-Service provides optimised, ready-to-deploy Visual Language Models tailored for Smart City applications.

End-to-end cloud infrastructure

To support the complete vision model lifecycle, Hafnia is built on a flexible, multi-cloud strategy that leverages AWS, Nebius, and other providers. By uniting foundational cloud reliability with specialised AI computing, Hafnia ensures that every stage of development has the exact power it needs. The Milestone Synthetic Data Generation pipeline using Cosmos Transfer and Cosmos Evaluator powered by Cosmos Reason was also launched on Nebius at GTC. Importantly, this multi-cloud approach supports data sovereignty requirements, ensuring customers maintain full control over where their sensitive information is stored and processed.

From initial data sourcing and customised fine-tuning, to full-scale training and deployment, these collaborations provide the scalable power needed to manage the complete lifecycle of vision model development.