On 2 September 2026, IBM and Confluent announced that IBM Granite Time Series models are available in Early Access on Confluent Cloud, bringing forecasting and anomaly detection directly to enterprise data streams. IBM had previewed the partnership on 1 September; IBM Research published the technical walkthrough on Hugging Face the following day. The integration pairs IBM's time-series foundation models with Confluent's streaming platform so teams analyze operational data as it is generated, calling models from Apache Flink SQL without standing up a separate machine-learning environment.

Early Access begins on Confluent Cloud on AWS, with Confluent Platform support for on-premises and hybrid environments planned next. The portfolio comprises four complementary models — PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse — all small by design (1M to 260M parameters, GPU-free) so inference can run inside the streaming pipeline. Users switch models with a single SQL parameter in the existing AI_FORECAST and AI_DETECT_ANOMALIES Flink functions, avoiding pipeline redesign.

What's new

  • PatchTST-FM-r1 — all-around performer, strongest at probabilistic forecasting with full distributions and quantiles.
  • FlowState-r1.1 — sampling-rate invariant, best point-forecast accuracy across SCADA and market data rates.
  • TTM-r3 — best efficiency/performance trade-off, supports control variables, runs hundreds of thousands of series nightly on CPU.
  • TSPulse — 1M-parameter multi-task model for anomaly detection, classification, similarity search, and gap-filling.

All four models are accessible through Confluent's native inference in Apache Flink on Confluent Cloud. Inference results are written to Kafka topics and shared with fanout, consumable by alerting systems, dashboards, lakehouses, and AI agents. Confluent manages model serving, infrastructure, scaling, and runtime operations — no provider credentials or glue between data pipelines and the model.

Why it matters

Enterprises have traditionally built and maintained a separate model for each dataset or application, limiting modeling to the few hundred series where the money is and covering the rest with safety margins. A single time-series foundation model generalizes to unseen series, letting a demand planner, fraud analyst, or process engineer put models to work on their own streams. IBM reports that every point of accuracy is worth millions in tested verticals (cement, steel, pulp and paper, food, telecommunications) and productivity gains run 5 to 10×, with work that waited for specialists now sitting with domain experts who own the decision.

The integration also brings IBM's enterprise AI governance framework — model provenance, licensing transparency, and built-in schemas, lineage, and access controls — to streaming inference. Kafka topics are durable and replayable, supporting auditing, troubleshooting, model evaluation, and rerunning inference against historical data.

Our take

The real shift is not just model quality but where inference lives: inside the streaming runtime, keyed per series, with Flink managing state fault-tolerantly. That eliminates the extract-transform-load latency that makes batch forecasts stale by the time they reach a decision. If the portfolio delivers on zero-config switching and CPU-only inference at scale, it removes the main reason teams keep safety stock instead of forecasting the long tail.

Sources