Python and Java SDKs added a securable type letting Databricks Apps reference Model Serving endpoints

The main product change in October 2026 was about connecting Databricks Apps to Model Serving endpoints. On October 10, the Python SDK (v0.151.0) and the Java SDK (v0.165.0) each added a new enum value for Unity Catalog securable types under Databricks Apps. Developers can use it to configure model serving endpoints as securable resources in app manifests [2][3]. A few days earlier, the v1.137.0 release expanded its resource exporter to cover model serving endpoints and Databricks Apps, along with Lakeflow pipeline settings, cross-cloud cluster policies and several new job task types [6].

Two published case studies showed Model Serving as the final production step. TinyFish described a path from post-training to deployment. It used Databricks AI Runtime on managed GPUs, MLflow for evaluation and Model Serving for production inference. Its web-native model reached a 77.6% task-success rate across 2,178 web-agent tasks in internal go/no-go benchmarking [4]. A separate reference architecture for e-commerce recommendations used AI Search for candidate retrieval, Lakebase for online feature serving and Model Serving for low-latency inference. It combined batch precomputation and real-time scoring on one governed platform [7].

Community discussion centered on choosing where to host inference logic. One post compared Databricks Apps and Model Serving on authentication, cost and performance, which suggests practitioners were weighing the two as deployment targets [1]. The new securable type fits this picture, since an app can now declare a serving endpoint as a governed resource rather than treating it as an external dependency [2][3]. Another community post covered using Genie Code outside the workspace [5].