Databricks AI Runtime and NVIDIA make it easier to customize open models with enterprise data

AI Runtime (AIR) integrates NVIDIA’s AI software and serverless GPUs with Databricks’ governance and data tools, allowing teams to fine-tune open models without managing GPU infrastructure. Developers can use existing code, containers, and frameworks like NVIDIA NeMo to experiment and scale training directly within Databricks. The platform handles underlying clusters and networking, enabling focus on model adaptation rather than infrastructure setup.

Unity Gateway provides governed access to frontier and open-source models, while AIR supports multi-node distributed training with NVIDIA Quantum InfiniBand and NCCL for high-performance communication. Teams can track experiments, GPU utilization, and data access through MLflow and Unity Catalog, ensuring reproducibility and governance throughout the workflow.

AIR now supports Docker and private registries, enabling developers to bring custom containerized environments to GPU workloads. Environment caching and NVIDIA profiling tools reduce setup time and help identify performance bottlenecks, while MLflow tracks compute metrics to optimize training runs.

Customized models can be deployed for real-time inference via Databricks Model Serving, completing the workflow from enterprise data to production-ready AI agents. Nemotron 3.5 Lightning and Super VL models are available for immediate experimentation and adaptation on Databricks.