Announcement | What’s New in Databricks: Agents, GPUs, and Real-Time ML

[Tushar_Parekar]

Databricks is expanding its AI Platform with new capabilities for ML engineering, deep learning, and real-time ML, helping teams move faster from model development to production on the same platform.

What’s new

• Genie Code for ML: Genie Code now supports more of the ML lifecycle, helping teams with feature engineering, model training, evaluation, deployment, and production troubleshooting with context from Unity Catalog and MLflow.
• AI Runtime for GPU training: AI Runtime brings serverless GPU training to Databricks with on-demand A10 and H100 access, deep learning environments, Lakeflow orchestration, and built-in governance through Unity Catalog and MLflow.
• Better support for real-time ML: Databricks is also highlighting stronger support for real-time ML through Feature Store and Model Serving, including streaming features, online feature serving, and high-QPS serving for demanding low-latency workloads.

Databricks positions these updates as a way to reduce the manual work around building and running ML systems, so teams can spend less time on infrastructure and more time on training, serving, and improving models in production.

👉 Read the full post here

[GabFernandes]

Great updates! I've been using Genie Code heavily for ML engineering in Databricks and the productivity gain is real.

Just finished building a full RFM segmentation model for supplier performance analysis end-to-end inside a notebook — feature engineering with Feature Store, K-Means training with Optuna hyperparameter optimization, MLflow experiment tracking, model registration in Unity Catalog, and segment profiling — all with Genie Code helping at each step.

What impressed me most is how context-aware it is: it reads your existing tables, understands the schema, suggests code that actually fits your data, and catches issues before they become bugs. It also handles the Unity Catalog + MLflow integration patterns correctly out of the box, which is usually the trickiest part.

For anyone studying for the ML Engineer Associate cert or building production ML pipelines on Databricks, I'd strongly recommend trying it. It doesn't just write boilerplate — it understands the full ML lifecycle.