Databricks AI Runtime Goes GA, Adds NVIDIA Tools for Fine-Tuning Open Models

Databricks said its AI Runtime (AIR) is now generally available, giving developers serverless access to NVIDIA GPUs for training and fine-tuning open models on enterprise data without managing GPU infrastructure directly.

The company made the announcement at GTC Berlin as part of an expanded collaboration with NVIDIA aimed at making it easier for enterprises to customize open models with their own data.

What AIR now supports

Databricks says AIR supports NVIDIA NeMo and NGC containers, along with multi-node distributed training. According to the company, NVIDIA Quantum InfiniBand provides high-bandwidth, low-latency networking through remote direct memory access, while the NVIDIA Collective Communications Library handles communication between GPUs. Databricks says it manages the underlying clusters and networking on behalf of developers.

The company also says AIR now supports Docker and private registries, letting developers bring custom containerized environments to GPU workloads. Databricks says developers can pull GPU-optimized containers from the NVIDIA NGC Catalog into the Databricks Artifact Registry, and that environment caching reduces setup time as teams rerun experiments.

Tools for training and debugging

According to Databricks, developers can prepare Lakehouse data and experiment in GPU-powered notebooks or through the AIR command-line interface, using environments that include PyTorch and CUDA. The company says Lakeflow Jobs can orchestrate training runs alongside data preparation and evaluation, while MLflow tracks training experiments and GPU utilization. Unity Catalog governs access to training data, and Databricks says native system tables provide visibility into CPU and GPU usage.

Databricks says its Genie Code tool can help developers write and adapt training scripts in a notebook, including debugging model code, identifying GPU performance bottlenecks such as memory issues and communication overhead, and applying NVIDIA-specific guidance on CUDA, cuDNN, and NCCL.

Nemotron and Kumo model support

With the release of Nemotron 3.5 Lightning and Super VL, Databricks says customers can load, experiment with, adapt, and serve these models starting day one. The company says AIR supports model customization through NVIDIA NeMo and NeMo Automodel, and cites an example of fine-tuning a Nemotron model on historical support tickets to build a technical support agent.

Databricks also says AIR supports a validated fine-tuning workflow for NVIDIA Kumo Anomaly, an open research model for time-series anomaly detection, with MLflow tracking and Unity Catalog checkpoints. The company says this pairs with its Genie Ontology to support NVIDIA's Agent Knowledge Foundation vision.

Databricks says that once models are fine-tuned and evaluated, teams can deploy them to production through Databricks Model Serving, completing what the company describes as a workflow from model customization on serverless NVIDIA GPUs to real-time inference.

Hanlin Tang, CTO for Neural Networks at Databricks, said AI Runtime has unlocked model training and reinforcement learning on enterprise data that was traditionally out of reach for machine learning engineers. Pat Lee, VP of Strategic Partnerships at NVIDIA, said the companies are bringing open models like Nemotron and Kumo directly to governed enterprise data to help teams build specialized agents.

Source: Databricks Blog, "Databricks AI Runtime and NVIDIA make it easier to customize open models with enterprise data," published October 21, 2026.