JFrog ML Release Notes

Welcome to the JFrog ML Release Notes. Below you will find the details for the latest release, followed by an archive of previous releases.

Latest Release: December 2025

Features:

Platform UI and AI Catalog Integration Updates: Integration details and model testing snippets now use the unified <tenantId>.ml.jfrog.iodomain for correct and consistent integration.
Update Python Support in FrogML: The FrogML SDK now supports Python 3.12 and 3.13.

Fixes

Changed from project key to project name in detected model view.
Resolved the inability to block models from the registry.
Fixed [FrogML] model logging failing for non-admin tokens due to missing project key in repository configuration.
RBAC set approval now requires Read permissions, not just allow.
Reports no longer fail if a build was deleted but the job still exists.
Updated execution manager TTL to 6 hours.
Fixed an issue where the autoscale policy wasn't attached if an existing one was present.

Fixes

Updated logic to ignore '-cache' suffix from Xray to get the original repository name.
Ensured that if a repository has no projects associated, it returns with empty projects associated.
Fixed mismatch between actuator and REST controller causing failure to increase log level in model catalog.
Fixed registry view requests to only fetch allowed models instead of all models.
Fixed bug: Models detected in a repository not associated with a project are now considered unmanaged.

Features:

Deployment as a Service: Introduced a "one-click" experience for deploying open-source Large Language Models (LLMs) from Hugging Face.
JFrog ML on Azure: Added support for JFrog ML environments on Azure.
AI Gateway: General availability of the AI Gateway.

Fixes

Fixed GPU build failure on Azure Clusters due to CDI device injection errors.
Fixed a routing issue on AI catalog models.
Fixed an issue where the secret list was not shown in batch automation.
Resolved an issue where executions were stuck without getting alerts when Spark application status was PENDING RERUN.
Removed unnecessary cache from FrogML get_credentials.

Features:

FrogML SDK Migration: Removed Qwak terminology from services and code examples as part of releasing the new FrogML SDK and CLI.
Improved Model Card for Open Source: Redesigned open source model cards to be more informative and user-friendly, standardizing metadata like license, version, supported tasks, and showing performance metrics.

Fixes

Fixed promote to build failure when model ID differs from model name.
Fixed -Traffic split and shadow deployments do not work from the SDK.
Fixed promote-to-build failure in self-managed environments.
Fixed [jfrog-gateway] Upon revoke of a token, if missing token (not found), return not found.
Fixed [Model Version] Temporary tokens are not revoked correctly.
Fixed incorrect page display when filtering on "Live models".
Fixed Unity Catalog source failure on validation (on execution OK).

Features:

LLM Model Deployment: Features for LLM model deployment, including deployment flow design, charts, and images definition.

Fixes

Running build with the --deploy flag failure on Qwak SDK and FrogML CLI fixed.
Username and password not working with Frogml CLI fixed.
Fixed AttributeError in FrogMLClient methods.
Made 'tags' and 'filters' optional in FrogMLClient.list_builds method.
Data Ingestor: Raise handled error for datasource validation in featureset validation.
Make CI for dependencies of fs-spark-addons to run CI of fs-spark-addons.
Fixed wrong build redirect URL in execution list of batch model.
Fixed the issue with environment not listed in token on the self-managed resolving in Builds Orchestrator.
Fixed use of only one cloud credentials / mount in Fitness builds.
Prevent project deletion if models exist.
Fixed FrogML build proto error.
Fixed batch execution logs filter on pagination.
Fixed ModuleNotFoundError in Qwak-SDK dev version.

Fixes

FrogML: Fixed validation error when passing None to the version field in Pydantic Model. The version field now defaults to the current datetime string if None is passed.
Hybrid Clusters: Restored the call to kube captain to apply model traffic for hybrid clusters.