IBM Granite Time Series models now available for real-time streaming inference on Confluent Cloud

IBM and Confluent have integrated IBM’s Granite Time Series foundation models with Confluent Cloud for real-time inference, launching in Early Access on AWS. The models are designed for forecasting, anomaly detection, optimization, and semantic intelligence, and are intended to run where business data already flows—stream-native and without separate ML infrastructure.

IBM describes its Granite Time Series models as trained once across vast, varied signals, capable of generalizing to unseen series. The company reports that internal deployments across industries such as cement, steel, pulp and paper, food, and telecommunications delivered productivity gains of 5 to 10× and accuracy improvements where “every point of accuracy is worth millions.”

The integration enables users to call Granite Time Series models directly from Flink SQL within Confluent Cloud, supporting real-time anomaly detection and forecasting without extracting data into a separate ML platform or data warehouse. Confluent manages model serving, infrastructure, scaling, and runtime operations, aiming for zero configuration and faster time to value.

Confluent Cloud provides native inference, allowing models to run directly within Apache Flink on Confluent. This unifies data and ML workflows, eliminates the need to provision dedicated model-serving infrastructure or GPUs, and avoids cloud ingress or egress fees. Inference results are written to Kafka topics and shared downstream to alerting systems, dashboards, lakehouses, and AI agents.

The platform emphasizes built-in governance, traceability, and security, with inference pipelines adhering to the same schemas, lineage, and access controls as the rest of the platform. Kafka topics are durable and replayable, supporting auditing, troubleshooting, model evaluation, and historical inference reruns.

Confluent Platform will follow, bringing the same models and capabilities to on-premises and hybrid environments. IBM states that the models have accumulated 44M+ downloads prior to this integration.