Databricks releases, every cloud
- Genie Code for AI Runtime (Public Preview)
Generates distributed training code and resolves environment issues. Requires AI Runtime.
- Web terminal on serverless GPU compute (AI Runtime) is in Public Preview
Runs shell commands, monitors GPU usage, and manages files on serverless GPU compute. Requires environment version 5 or above.
- Protobuf tensor input for custom model serving endpoints (Public Preview)
Replaces JSON with serialized KServe v2 ModelInferRequest. Requires endpoints deployed after July 9, 2026.
AWS, read it on their docsAzure, read it on their docsGCP, read it on their docsSAP(not on this cloud) - Custom Docker images for AI Runtime CLI workloads (Beta)
Allows custom system library versions and complex dependencies, requires setup via Use custom Docker images.
- AI Runtime CLI (Beta)
Submits and manages distributed training workloads on serverless GPU compute from a local machine.
- Serve custom LLMs with Custom Model Serving (Beta)
Supports multimodal models and PEFT recipes, unlike Foundation Model APIs.
- Run the change, do not just read it. Hands-on labs in your own Databricks workspace, graded when you submit. Browse labs
- Alibaba Cloud Qwen3.5 122B A10B now available as a Databricks-hosted model in Public Preview
Accessible via Foundation Model APIs, available in Public Preview.
- AI Runtime 1xH100 accelerator (Beta)
Supports 1xH100 accelerator, see Hardware options for details.
- AI Runtime is now in Public Preview
Adds GPU support to serverless compute for deep learning workloads.
- Endpoint telemetry for custom model serving endpoints (Beta)
Stores logs, traces, and metrics in Unity Catalog Delta tables using OpenTelemetry.
- Alibaba Cloud Qwen3-Next Instruct now available in Beta
Supports 80B A3B Instruct in us-west-2 and ap-northeast-1 regions via pay-per-token APIs.
- Improved autoscaling behavior for Mosaic AI Model Serving
Ignores brief traffic surges, scaling on sustained load increases. Reduces unnecessary concurrency scaling.
- Anthropic Claude Opus 4.1 now available as a Databricks-hosted foundation model
Accessible via Foundation Model APIs on a pay-per-token basis.
- Route-optimized endpoints now require route-optimized URL path for querying
Requires route-optimized URL path for querying, standard workspace URL path not supported for new endpoints.
Headlines, dates and product areas are Databricks' own, from the release notes published for each cloud, and every item links to the note it came from. The one-line summaries are ours. Not a Databricks product and not affiliated with Databricks, Inc.