IBM and Confluent Launch Streamhouse Architecture for Native Time-Series Foundation Models in Apache Flink
·2026.11.19 00:00
Key point
The Streamhouse architecture enables CPU-based time-series forecasting and anomaly detection within streaming pipelines, with models downloaded over 50 million times.
Details
- IBM Research and Confluent announced the 'Streamhouse' architecture on November 18, 2026, integrating time-series foundation models natively into Apache Flink streaming data.
- The system runs on standard CPUs without requiring GPUs, reducing infrastructure complexity and deployment costs for real-time AI applications.
- Models have been downloaded over 50 million times and are available via Confluent Cloud (fully managed) or on-premises via Confluent Platform for Apache Flink.
- Key use cases include production optimization, anomaly detection, supply chain planning, fraud and security, trading, and semantic search, all leveraging pre-trained models to update predictions as new data arrives.
- The architecture supports industry-specific workflows in industrial AI, retail, banking, and airlines by combining Flink's processing with Granite models for forecasting, detection, and optimization.
- Users can access code and notebooks at github.com/ibm-granite/granite-tsfm and utilize functions like AI_FORECAST and AI_DETECT_ANOMALIES within SQL queries.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.