Cloudera Docs

What's New

Major features and updates for the Cloudera AI data service.

February 19, 2026

Release notes and fixed issues for version 2.0.55-b193.

New Features / Improvements

Cloudera AI Workbench
  • Improved application responsiveness by optimizing database query performance across projects and jobs, significantly reducing page load times for environments with large datasets.
Cloudera AI Control Plane
  • Improved consistency across the administrative experience for AI Workbench, AI Registry, and the AI Inference services.
  • Added support for Amazon EKS 1.33.
  • Added support for Azure AKS 1.33.
  • Improved the reliability of Cloudera AI Inference service upgrade processes.
  • Cloudera AI now automatically performs side-by-side upgrades for incompatible environments, ensuring a seamless one-click experience with built-in rollback support in case of failure. To upgrade a workbench, Cloudera users must hold both the MLAdmin and EnvironmentAdmin roles, and those performing an upgrade must also possess the necessary cloud provider permissions to execute backup and restore operations for the underlying storage and metadata databases. For more information, see Upgrading Cloudera AI Workbenches.
Cloudera AI Inference service
  • Cloudera AI Inference service now provides a production-grade serving environment for hosting applications. Applications deployed on Cloudera AI Inference service can scale alongside Model Endpoints, providing a scalable solution for various components. For more information, see Serving Applications on Cloudera AI Inference service (Technical Preview) .
  • Cloudera AI Inference service now supports AWS on-demand capacity reservations and capacity blocks to ensure compute availability for inference workloads. For more information, see Configuring AWS on-demand capacity reservations and capacity blocks.
  • Cloudera AI now supports the deployment of Hugging Face reranking models using the API.
  • Cloudera AI now supports deploying Hugging Face embedding models using the API.
  • You can now manually specify model tasks (such as EMBED, RANK, or CLASSIFICATION) using API during deployment, enabling broader vLLM support for architectures like bertmodel or modernbertfortokenclassification that serve tasks like embedding and reranking respectively.
  • You can now manage Cloudera AI Inference service logging globally using the Serving API ConfigMap, allowing administrators to enable logging and define a storage bucket across all endpoints simultaneously for consistent data collection.
  • Added Fine Grained Authorization support. For more information, see Configuring Fine-grained Access Control.
Cloudera AI Registry
  • Cloudera AI Inference service now supports direct deployment for XGBoost, PyTorch, and TensorFlow models using the AI Registry. For more information, see Deploying Additional Model Frameworks.
  • Cloudera AI Registry now displays structured metadata and comprehensive lineage tracking (provider, model ID, and SHA) for all models imported from Hugging Face and NVIDIA NGC.

Fixed Issues

Cloudera AI Workbench
  • Resolved an issue that prevented Team Administrators from updating the team description in the UI. (DSE-48770)
  • Fixed the workload status synchronisation issue between dashboards and dashboard_pods, ensuring accurate and consistent resource usage reporting. (DSE-46977)
Cloudera AI Inference service
  • Resolved an issue where Hugging Face model IDs in inference URLs were un-sanitized, causing routing failures; inference endpoints now correctly use underscores (_) instead of slashes to align with Triton directory structures and authorizer patterns. (DSE-50543)
  • Fixed an issue where Hugging Face embedding models failed to deploy to vLLM due to an unsupported embed task key. (DSE-49993)

October 31, 2025

Release notes and fixed issues for version 2.0.53-b241.

New Features / Improvements

Model Hub
  • Improved model information display by surfacing attributes like MaxTokens, Parameters, and Dimensions for Embedding Models, enabling better decision-making before importing models.
Cloudera AI Inference service
  • Added the ability for users to start and stop deployed model endpoints, providing greater control over resource management and cost optimization. This feature allows you to pause inactive endpoints to save resources while maintaining the ability to quickly restart them when needed.
  • Improved user experience by enabling customers to open model endpoints in different browser tabs, allowing for better multitasking and simultaneous monitoring of multiple endpoints.
  • Enhanced API accessibility by providing customers access to the Swagger UI for invoking AI Inference APIs, enabling interactive API testing and documentation exploration.
  • Added visibility to the underlying compute cluster through a direct link in the UI for AI Inference instances, providing seamless navigation to cluster details and infrastructure monitoring.
  • Implemented validation for root volume size when adding new nodes to AI Inference deployments, preventing configuration errors and ensuring adequate storage capacity.
  • Enhanced node group management by displaying GPU types and enabling filtering based on GPU type when searching through available nodes, improving resource selection efficiency.
  • Upgraded KServe from version 0.12 to 0.15, enhancing the underlying model serving infrastructure.
Cloudera AI Platform
  • Enhanced support for self-signed certificates in public cloud deployments, resolving installation and update pain points.
  • Re-enabled cadence workflow for workbench upgrades, fixing version compatibility issues.
Cloudera AI Workbench
  • Optimized the loading speed of the Site Administration overview and the Projects pages by improving cluster-wide resource usage data collection, ensuring quick loading even in environments with 1,000+ user namespaces.
  • The Job Retry feature introduces automated recovery for failed, timed out, or skipped jobs. Administrators can now set concurrent retry limits and customize the retry behavior to enhance job resilience and eliminate the need for manual failure intervention.

Fixed Issues

Cloudera AI Registry
  • Fixed Events & Logs rendering during the creation of new registries in the UI. (DSE-48101)
Cloudera AI Inference service
  • Resolved inconsistencies in the tooltips displayed on the Model Endpoint Details page, ensuring accurate and helpful information is shown to users. (DSE-45042)
  • The API service can now inspect the metadata of MLflow models from the Cloudera AI Registry to correctly identify the Hugging Face transformer flavor. This ensures the model is automatically routed to the correct runtime (i.e. HuggingFace runtime with vLLM backend unlike defaulting to Triton + ONNX earlier), enabling successful deployment of MLflow transformer models. (DSE-46462)
Cloudera AI Workbench
  • Removed the rigid requirement for every team to have an Administrator, allowing Team Administrators (or Site Administrators if none are designated) to manage team membership and roles. (DSE-41700)
  • You can now customize the model build process with two new options: model_root_dir allows setting a custom source directory, and build_script_path enables specifying a custom location for the build script. (DSE-45166)
  • Resolved a page layout issue that occurred after horizontal window resizing. The fix ensures the browser view restores correctly without requiring a manual refresh. (DSE-40393).
MLRuntimes
  • Fixed an issue where system messages were hidden or suppressed when a Workload (such as a Session, Job, or Application) failed to start up while using a PBJ Workbench Runtime. (DSE-46126)

Behavioral Change

MLRuntimes
  • The R kernel output mechanism has been updated to improve visibility of interactive commands. The output of R commands (for example, install.packages("sparklyr")) is now mirrored to both the Session tab and the Logs tab by default, rather than appearing solely in the Logs. This ensures immediate visibility of important messages like installation progress, compilation output, and errors. Users can revert to the previous behavior by setting the environment variable JUPYTER_KERNEL_OUTPUT_FILTER_REGEX to "DISABLED" or "^$". (DSE-35335)
  • Clearing a session now permanently removes all content from the session history, ensuring that subsequent PDF exports and email reports contain only current, relevant results and no stale data from previous runs. (DSE-12565)
Cloudera AI Inference service
  • MLFlow integration with transformer models is now complete, enabling models saved using the workbench (via MLFlow) in the Hugging Face transformer format to be successfully deployed on Cloudera AI Inference service, including both regular and finetuned models. (DSE-43840)

September 25, 2025

Release notes and fixed issues for version 2.0.52-b60.

New Features / Improvements

Cloudera AI Registry
  • You can now update the visibility of models directly from the UI.
  • You can now delete registries in a failed status directly from the UI.
  • Added the ability to use pre-downloaded artifacts when importing external models.
  • The model card for embedding models now includes MaxTokens, Parameters, and Dimensions.
  • Added support for upgrading Azure registries with UDR-enabled subnets.

  • Improved error resolution suggestions have been added to the UI for Cloudera AI Registry, empowering users to troubleshoot issues on their own.

Cloudera AI Inference service
  • Added support for new models, including NeMo Retriever’s GraphicElements, PageElements, and TableStructure, as well as PaddleOCR, Boltz-2, and GPT-OSS models.
  • The Test Model feature is now available for Speech-to-Text models.
  • You can now specify vLLM arguments when deploying Hugging Face models.
  • Improved error resolution suggestions have been added to the UI for Cloudera AI Inference service empowering users to troubleshoot issues on their own.
  • Grouped CPU/GPU nodes under the Cloudera AI Inference service Details page for a more user-friendly experience.
Cloudera AI Platform
  • You can now provision multiple CPU/GPU resource groups in the Cloudera AI Workbench. This provides enhanced control over workload scheduling and allows segregation of workloads based on the instance types.

  • Added support for EKS 1.32.

  • Added support for AKS 1.32.

Fixed Issues

Cloudera AI Inference service
  • Addressed critical CVEs in most NGC models by upgrading to the latest NIM versions. (DSE-47435)

  • Resolved an issue where code samples were not correctly rendered for riva, reranking, and retrieval models. (DSE-46413)

  • Ensured that the UI carries out proper validation of GPU, CPU, and memory during model deployment. (DSE-46250)

  • Resolved an issue where Prometheus errors were not surfaced in the Model Endpoints UI. (DSE-46077)

  • Addressed an accessibility issue with the 'Deploy external dropdown button'. (DSE-46071)

  • Resolved an issue that blocked navigation to other tabs while waiting for a response from the Model Endpoint's Test Model. (DSE-46247)

  • Resolved an issue where a scale alert box was not displayed in the UI when the model endpoint was ready and the current replica count was zero. (DSE-46301)

Cloudera AI Registry
  • Resolved an issue where the UI did not properly surface failures occurring during NVIDIA model imports from the Model Hub page. (DSE-45971)
  • Resolved an issue where the UI was not correctly auto-selecting the running registry. (DSE-46246)
  • Resolved an issue that prevented users from registering models from within a workbench. (DSE-47077)
Cloudera AI Workbench
  • Resolved an issue where proxy environment variables were being overwritten during model builds. (DSE-46070)
  • Resolved an issue where the Project Settings page displayed the project owner's ID instead of their name. (DSE-46572)
Cloudera AI Platform
  • Resolved an issue where spaces were incorrectly added to the ldap_dn during synchronization of some users, which caused those members of CML groups to be unintentionally removed. (DSE-47591)