Integration of IBM Time-Series Models
On November 18, 2026, IBM and Confluent announced the integration of IBM time-series foundation models into Confluent’s Streamhouse architecture. This family of four models has achieved over 50 million downloads and is now available as built-in functions in Confluent Cloud and on-premises via Confluent Platform for Apache Flink.
Deployment and Architecture
The time-series models operate on standard CPUs, eliminating the need for GPUs. Apache Flink is utilized to prepare and enrich streaming data, while Lightning Tables ensure low-latency access to the current state of data. The Streamhouse is designed as a continuous pipeline that connects, streams, processes, governs, and queries live business context, making model outputs immediately accessible to applications and agents.
Models can be invoked using SQL within Flink, and Confluent Cloud provides a fully managed option that requires no deployment. For on-premises users, the same models can be executed within their own infrastructure using Confluent Platform for Apache Flink.
Capabilities and Example Workflows
The time-series foundation models serve as pretrained starting points for various tasks, including forecasting, anomaly detection, and semantic intelligence over numerical streams. Key components include AI_FORECAST, AI_DETECT_ANOMALIES, TSPulse, and PatchTST-FM, with the Granite models family being referenced for forecasting and embeddings.
Industry Use Cases
Several industry scenarios illustrate the application of these models:
- Manufacturing: A pretrained model analyzes SCADA telemetry and QA measurements to forecast product quality hours in advance and suggest optimal control setpoints.
- Retail: The model generates refreshable 28-day SKU forecasts that inform reorder calculations, ensuring stock availability without excess inventory.
- Fraud Detection: The models assess individual cardholders’ expected spending ranges, providing additional context to existing fraud detection systems before transactions are approved.
- Trading: PatchTST-FM forecasts short-term trading volume, assisting execution desks in pacing large orders effectively.
- Semantic Search: TSPulse enables similarity searches across time-series windows, aiding reliability engineers in identifying patterns in sensor data.
IBM and Confluent have made code and notebooks available at github.com/ibm-granite/granite-tsfm.
Original source: Hugging Face Blog