IBM and Confluent are bringing together time series foundation models (TSFM) for use in real-time data streams. While the traditional approach requires developing a separate model for each data series and months of work by specialists, these models can generalize to series they have not seen before, enabling forecasting, anomaly detection, similarity search, classification, missing data imputation and optimization. This is intended to reduce the margins allocated for safety stocks, excess capacity and decisions made too late.
In the solution, IBM Granite Time Series models analyze the signals, while Confluent delivers the business’s live data as a stream. Through native inference on Apache Flink, the models can run directly within Confluent Cloud. Flink maintains the historical data required for each series in a fault-tolerant manner, enabling forecasting and anomaly detection based on up-to-date context. The service is being made available on Confluent Cloud for AWS under Early Access; it is also planned to be extended to on-premises and hybrid environments through Confluent Platform.
IBM and Confluent are offering four complementary models under Early Access for different use cases. PatchTST-FM processes variables in separate channels to produce a probability distribution. FlowState tracks continuous-time dynamics, from second-by-second SCADA data to hourly market data. TTM offers a compact architecture capable of processing 100,000 series overnight on a CPU. TSPulse focuses on anomaly detection, classification, missing data imputation and searching for similar events in the past. The models can be used with Confluent’s AI_FORECAST and AI_DETECT_ANOMALIES Flink SQL functions by changing just one SQL parameter.
Application areas include retail demand planning, financial fraud detection and manufacturing optimization. For new products, the models can leverage similar series without historical data; in payment flows, they can identify deviations from a card’s normal behavior before the transaction is completed. In manufacturing, they can recommend new settings based on variables such as temperature, speed, dosage and viscosity, in line with targeted performance and constraints.
IBM says it tested the solution first in its own products and operations, and then with design partners in the cement, steel, paper, food and telecommunications sectors. According to the company, efficiency gains can reach 5 to 10×. The Granite models also use IBM’s enterprise AI governance framework, which includes model provenance and licensing transparency.
Why it matters
This integration aims to transform time-series analytics from projects in which data science teams build models separately into a service that can be used within the platform where live data streams reside. This could enable operational decisions, particularly in retail, finance and manufacturing, to be based not only on historical reports but also on real-time signals and similar events; however, accuracy, the false alarm rate and how the recommended settings will be adapted to business processes remain open questions. While running the models within Confluent Cloud could lower the barrier to implementation for businesses using existing Flink and SQL infrastructure, extending the solution to on-premises and hybrid environments is still in the planning stage. Access and usage are limited due to its Early Access status; whether the efficiency gains cited by IBM can be replicated across different sectors, and under what conditions, will also depend on how the tests progress.
Background
IBM is not a new name in the FikirPilot archive: in the past 90 days, we published a news story in which this name appeared; that article is dated August 24, 2026.