IBM time-series models now run natively in Confluent
IBM Research says its time-series foundation models now run as built-in functions in Confluent, in the cloud and on premises. The models use CPUs and aim to bring forecasting and anomaly detection closer to streaming business data.

Key takeaways
- IBM says its time-series models now run as built-in functions in Confluent.
- The models can run in Confluent Cloud or on premises using CPUs.
- IBM describes possible uses in manufacturing, retail, banking and airlines.
- The post presents examples, not independent benchmarks or named customer results.
In this article
What changed
IBM time-series models are now available natively inside Confluent, according to IBM Research. The announcement appeared in a blog post on Hugging Face on 18 November 2026. The post says the models can run in Confluent Cloud and on premises. They use CPUs, standard processors, rather than GPUs, specialised chips often used for AI work.
The integration makes the models available as built-in functions. Teams can call them from SQL, a common language for querying data. Data moves through Apache Flink, a stream-processing engine. In Confluent Cloud, IBM says the models are fully managed. On premises, they can run within a customer's own infrastructure using Confluent Platform for Apache Flink.
Why time-series models matter

Time-series data is a record of change over time, such as machine readings, payments, sales, orders or inventory levels. IBM says such data can help teams anticipate what happens next. The result must arrive while a decision can still be changed.
The post describes time-series foundation models as systems trained on broad historical data and then adapted to new tasks. Instead of building a separate model for every machine, product or store, teams can start from a shared model. IBM says teams can also fine-tune a model. They can adjust it with extra data or combine its outputs with business rules.
IBM says the model family has more than 50 million downloads. The post refers to four models behind that figure, but it does not give independent benchmark results.
Streamhouse and real-time context
IBM and Confluent frame the work around a Streamhouse. This setup continuously captures, processes, governs and serves current business data. Confluent's Data Streaming Platform combines five capabilities: connecting, streaming, processing, governing and querying. Apache Flink can change and enrich data while it moves. Lightning Tables then provide access to the current state with little delay.
According to the post, these models bring prediction into the architecture. Flink can prepare data and call forecasting and anomaly detection. Anomaly detection flags unusual behaviour. Flink can put results beside the events that caused them. Other kinds of models can also contribute. These include large language models, statistical tools and specialised machine learning systems.
Use cases and cautions
The post gives examples rather than confirmed customer results. In one scenario, an ice cream production line uses an already trained model. It ingests SCADA telemetry and quality measurements every 10 minutes. SCADA is a control system used in industrial plants. The model predicts the proportion of tubs likely to meet specifications four hours ahead, along with a likely range.
If the forecast falls below target, an optimisation routine tests different combinations of freezer temperature and airflow. The operator approves suggested settings, while routine changes can be automated within set limits. IBM says similar methods could be used for ovens, dryers, kilns and mills.
In another example, a retail team watches sales for thousands of products across hundreds of stores. The post says fixed limits may flag busy weekends but miss smaller shifts. A function named AI_DETECT_ANOMALIES creates an expected range for each sales series using recent history. It checks new readings against that range at a chosen confidence level. It does this without building a distinct model for each product and store.
What it means for teams
For teams already using Confluent or Apache Flink, the main appeal is location. Forecasts and anomaly signals can be moved nearer to operational applications, where people or automated systems can act on them. IBM lists possible uses in production quality, store activity, product planning, card payment checks and trading volume forecasts.
IBM says the components could be used in manufacturing, retail, banking and airlines. The models can run on standard CPUs next to the streaming pipeline. IBM says this may reduce the need for dedicated GPU hardware.
The claims come from IBM Research, and the post does not provide independent test results. The scenarios are presented as examples, not as proof of results. Organisations should check Confluent documentation and their own requirements before deployment.
Sources
- Time-Series Models in the Streamhouse: A New Frontier and Use CasesHugging Face Blog
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



