# IBM time-series foundation models integrate with Apache Flink on Confluent for CPU-based stream processing

> **Key takeaway:** IBM's pretrained time-series foundation models, downloaded over 50 million times, now run natively as built-in functions inside Confluent's streaming platform using Apache Flink and standard CPUs without GPU requirements.

**Published:** 2026-11-18T15:00:23+00:00 
**Source:** Hugging Face Blog &amp; Announcements 
**Category:** ai-open-models 
**Canonical URL:** https://fosswire.org/news/ibm-time-series-foundation-models-integrate-with-apache-flin.html 

## Summary
IBM Research and Confluent have integrated IBM's family of time-series foundation models directly into Confluent's data streaming platform, enabling real-time forecasting, anomaly detection, and semantic intelligence on streaming data through SQL queries. The models, which have accumulated more than 50 million downloads, are available as fully managed functions in Confluent Cloud and as deployable components in Confluent Platform for Apache Flink on premises. Unlike typical AI workloads, these compact models run on standard CPUs rather than GPUs, allowing them to operate alongside streaming pipelines without dedicated accelerator infrastructure.

The integration targets what the companies term a 'Streamhouse' architecture, where Apache Flink prepares and enriches data as it flows, Lightning Tables provide low-latency state access, and time-series models add predictive capabilities alongside large language models and statistical systems. Use cases span manufacturing quality control, retail anomaly detection, payment risk assessment, and trading volume forecasting, with the models serving as pretrained starting points that teams can fine-tune or combine with business logic rather than building separate models for each time series.

## Architecture and deployment model
- **CPU-only inference:** The time-series models are compact enough to run on standard CPUs, eliminating GPU infrastructure requirements and simplifying deployment alongside existing streaming pipelines in both cloud and on-premises environments.
- **Apache Flink SQL integration:** Models are exposed as built-in functions callable through SQL within Apache Flink, allowing teams to embed forecasting and anomaly detection directly into stream processing logic without separate model serving infrastructure.
- **Four-model family with 50M+ downloads:** IBM's time-series foundation model family has achieved substantial distribution through Hugging Face, suggesting broad community validation of the pretrained weights for general time-series pattern learning.
- **Streamhouse architecture layering:** The deployment positions time-series models within a broader intelligence stack alongside LLMs, statistical models, and specialized ML systems, with Flink orchestrating data preparation and result routing to operational applications.
- **Dual deployment modes:** Confluent Cloud offers fully managed execution with no deployment overhead, while Confluent Platform for Apache Flink supports on-premises and self-managed infrastructure for regulated or latency-sensitive environments.

## Why it matters
For the open-source AI and data infrastructure ecosystem, this integration demonstrates a viable path for foundation model deployment that avoids GPU dependency, potentially expanding access to real-time ML for organizations without accelerator budgets. The tight coupling with Apache Flink, an established open-source stream processing engine, rather than proprietary IBM infrastructure, suggests the approach could be replicated or extended by other model providers seeking streaming deployment patterns.