DEV Community

Mikhail Savchenko
Mikhail Savchenko

Posted on Originally published at inite.ai

IBM time-series models land natively in Confluent streaming

IBM Research announced on Hugging Face on November 18, 2026 that its time-series foundation models now run natively inside Confluent. They are exposed as built-in functions, so the work happens where the data already flows.

What is available

In Confluent Cloud the models are fully managed and need no deployment. On premises, they run inside the customer's own infrastructure through Confluent Platform for Apache Flink. IBM says the models are compact enough for standard CPUs and do not require GPUs. The post links to documentation for both routes.

With Apache Flink, teams apply forecasting, anomaly detection and what IBM calls semantic intelligence to streaming data using SQL. The post names two functions: AI_FORECAST and AI_DETECT_ANOMALIES. IBM says the models have passed 50 million downloads and belong to a family of four. The post names TTM, PatchTST-FM and TSPulse, and mentions the Granite models in its summary.

The pitch: one pretrained starting point

Time-series foundation models learn general patterns from varied historical data and are then applied to new series. IBM's argument is that this gives a common starting point across many series, so teams need not build and maintain a model for each one. Teams can also fine-tune a model or combine its output with other data and business logic. Forecasts can take in recent observations as data arrives, subject to the model's requirements and the pipeline's setup.

Scenarios IBM describes

The post lays out six scenarios. IBM frames them as illustrations of how organizations could combine the pieces, not as reported customer deployments, and the figures are simulated.

  • Retail anomaly detection: each product-and-store sales series gets an expected range, so quiet drops are flagged rather than only weekend peaks.
  • Supply planning: a 28-day demand forecast with a P10-P90 range feeds a reorder point, and an agent drafts a purchase order for human review.
  • Fraud signals: a payment outside a card's own expected spending range goes to the bank's existing fraud systems as one more input.
  • Manufacturing: a model forecasts the share of product in spec four hours ahead and helps evaluate setpoints before an operator approves a change.
  • Trading: volume forecasts help pace large orders.
  • Similarity search: embeddings of sensor windows support searching for similar patterns, though this needs a separately built embedding pipeline and index.

IBM states plainly that similarity search can narrow an investigation but does not establish a root cause by itself.

What is not in the post

The post gives no pricing, no benchmark comparison and no list of customers. Whether the quality of forecasts holds up on a given company's data is something each team will have to test.

Top comments (0)