IBM Granite time-series models land on Confluent, turning forecasting into SQL

IBM's time-series models, from the Granite family, now run natively inside Confluent. IBM Research detailed the integration on Wednesday (18/11) in a post on the Hugging Face blog: forecasting, anomaly detection and optimisation become built-in Apache Flink functions, called directly from SQL — on ordinary CPUs, with no GPU and no separate model server.

The integration has been in early access on Confluent Cloud since September. What arrives now is the public deep dive into the architecture, the general availability of Managed Model Inference for LLMs and embeddings, and the open preview of the Granite models in Confluent Platform for Apache Flink, for use inside the customer's own infrastructure.

What changes in practice

Teams working with streaming data can now call the models like any other SQL function. The AI_FORECAST function predicts the next values of a series, and AI_DETECT_ANOMALIES flags patterns outside the expected range. Behind them are time-series foundation models, pretrained on varied historical data: they generate forecasts for series they have never seen, without the team having to train a separate model for each dataset — so-called zero-shot inference.

The available models are IBM Granite TTM, lightweight and efficient for forecasting many series at once; Granite FlowState, for irregular or mixed-frequency data; Granite PatchTST-FM, which makes probabilistic forecasts and expresses uncertainty; and Granite TSPulse, built for anomaly detection. Google TimesFM is also supported. The platform is model-agnostic: switching models means changing one parameter, without rewriting the query.

Managed Model Inference also covers LLMs and embeddings, now generally available — via AI_COMPLETE and AI_EMBEDDING — and text analysis (sensitive-data detection and sentiment) in early access. The LLMs come from the three major cloud providers: Claude on Amazon Bedrock, GPT-5.4 on Azure OpenAI, and Gemini on Google Vertex AI. For LLMs and embeddings, customers pay the cloud provider's token price, with no Confluent markup.

Forecasting without a GPU

The most striking point is the hardware: the Granite time-series models are compact enough to run on standard CPUs, alongside the streaming pipeline — in the cloud or on premises. Confluent manages the inference infrastructure, with no provider account, API key or endpoint to configure. In Confluent Cloud, the models are fully managed; on premises, they run inside the customer's infrastructure via Confluent Platform for Apache Flink, in open preview with three models (TTM-R3, FlowState and PatchTST-FM).

Early access to the forecasting and anomaly functions runs on Confluent Cloud on AWS, in the us-east-1, us-east-2 and us-west-1 regions, with more regions on the roadmap. In a year of GPU scarcity, the promise is straightforward: time-series intelligence with no accelerator queue and no model server to operate.

Six use cases — scenarios, not customers

IBM Research's post describes six illustrative scenarios for combining the functions with streaming data — not documented customer deployments. Among them: production optimisation, in which the TTM model reads telemetry from an ice-cream line and forecasts, four hours ahead, the share of tubs within specification; retail anomaly detection, comparing each product's sales with the expected range; supply planning, with a 28-day demand forecast and a P10–P90 confidence interval to calculate the reorder point; card fraud, with each transaction compared with the card's own spending pattern; trading, in which PatchTST-FM forecasts trading volume for the next 30 minutes; and semantic search, in which TSPulse produces representations of time-series windows to find similar patterns in other equipment.

What has not been disclosed yet

The announcement includes no accuracy, latency or throughput benchmarks, no pricing for the time-series functions, and no customer results. The figure of more than 50 million downloads for the four-model family comes from IBM, with no measurement period or comparison baseline disclosed. The numbers should be read for what they are: the promise of an architecture, not a result measured in production.

Context

Confluent is an IBM company, and the integration builds on the Streamhouse architecture, announced on 4 November, which organises the streaming platform into five capabilities — connect, stream, process, govern and query — to keep business context current for AI agents and applications. Geekiki already covered the Granite time-series family in October, when IBM brought four models together in a forecasting ensemble (read here).

Image: IBM logo, via Wikimedia Commons.

Sources: IBM Research no blog do Hugging Face; Confluent — Managed Model Inference; IBM Newsroom — arquitetura Streamhouse; Confluent — novidades do 4º trimestre de 2026; IBM Community — acesso antecipado de setembro; GitHub — IBM Granite TSFM.

Transparency: this article was produced with the assistance of artificial intelligence, based on public sources on the internet listed above. We recommend consulting the original sources to confirm the facts.

By Muse