Skip to main content
Back to timeline
Hugging Face - BlogSource publication:

IBM's time-series foundation models now run natively inside Confluent: SQL forecasting and anomaly detection on CPUs

Synopsis

IBM and Confluent have made IBM's time-series foundation models, downloaded more than 50 million times, built-in functions inside Confluent: fully managed in the cloud, deployable on premises, and callable from Apache Flink with SQL to forecast, detect anomalies and support similarity search as data flows. Previously such signals often became decisions only after batch processing or post-hoc lab checks, and teams frequently trained a separate model per series; a pretrained model can now serve as a common starting point and run on standard CPUs alongside the streaming pipeline. The material describes intended workflows for manufacturing, retail, banking and airlines, not measured deployment outcomes.

AI-generated editorial illustration: Time-Series Models in the Streamhouse: A New Frontier and Use Cases

Interpretation

Time-series foundation models now run natively inside Confluent as built-in functions for streaming data, fully managed with nothing to deploy in Confluent Cloud and running on the user's own infrastructure on premises through Confluent Platform for Apache Flink. Previously time-series signals often sat across systems that update at different times or relied on periodic batch processes, so a forecast could arrive after the production run, stockout or payment authorization had already happened. The change rests on the reach of the model family itself: the material states these IBM time-series models have more than 50 million downloads and run on CPUs without requiring GPUs.

The invocation pattern embeds the models as SQL functions in Flink's dataflow: Flink prepares the data, invokes forecasting and anomaly detection, and makes the results available alongside the events that produced them for applications and agents. A pretrained model offers a common starting point across many series, reducing the need to develop and maintain a separate model for each one, and teams can fine-tune or combine outputs with other data and business logic depending on the task. The material names the functions: AI_FORECAST for demand forecasting, AI_DETECT_ANOMALIES giving each sales series an expected range from its recent history, and TSPulse producing time-series window representations for similarity search.

Perspective

The material describes how time-series models, streaming data and application logic combine in manufacturing, retail, banking and airline scenarios, with figures labeled as simulated data; it addresses engineering teams that need to call forecasting and anomaly detection inside a streaming pipeline.

The material states that similarity search over time-series embeddings can narrow an investigation but does not establish a root cause by itself, and that whether forecasts incorporate recent observations depends on the model's requirements and the pipeline's configuration.