๐
November 19, 2026 โข Glonce / IBM Research on Hugging Face
IBM Research published the post "Time-Series Models in the Streamhouse" on the Hugging Face blog on November 18, 2026, written by IBM Research staff. The core announcement: IBM time-series foundation models now run natively inside Confluent as built-in functions, in the cloud and on premises โ part of the streamhouse idea of continuously capturing, processing, governing and serving current business context. The family spans four models with more than 50 million downloads per IBM, a figure with no stated time period, per Glonce.
The four models named across coverage are TTM (Tiny Time Mixer), PatchTST-FM, FlowState and TSPulse. They run on CPUs without requiring GPUs, small enough to sit next to the streaming pipeline and reduce dedicated GPU infrastructure. The positioning is a zero-shot starting point: a pretrained model applies to a new series, without a separate model per series, and fine-tuning stays optional. Model code is public in the ibm-granite/granite-tsfm GitHub repository. Per Glonce and Agentic Affairs, no benchmark, accuracy or cost figures were published โ strength claims rest on IBM descriptions.
How it works in practice. Two Flink SQL entry points are documented: AI_FORECAST predicts future values with a range, and AI_DETECT_ANOMALIES builds an expected range from recent history, flagging readings outside it at a chosen confidence level. Per Confluent documentation, cloud AI_FORECAST is Early Access for evaluation and non-production testing without support or an SLA, and on-premises AI functions are preview features whose interfaces may change, while on-prem model servers use plain HTTP without authentication, so Confluent docs recommend a Kubernetes NetworkPolicy.
Per independent reviews, these are illustrative scenarios, not reported deployments: production line โ pretrained TTM reads SCADA telemetry and QA measurements every 10 minutes and forecasts the share of in-spec tubs four hours ahead, retail โ a 28-day demand forecast with a P10โP90 range refreshed as sales arrive, store anomaly and card-payment fraud signals, trading โ PatchTST-FM forecasts the next 30 minutes on one-minute bars with a P10โP90 range, and TSPulse adds similarity search over historical patterns. Per ZGLG, closer to a reference architecture than to production proof.
Sources:
huggingface.co/blog/ibm-research/time-series-powered-streamhouse
glonce.com/ibm-brings-time-series-models-to-confluent-flink
trendcueai.com/news/aef632472a5e3af7f91b
agentic-affairs.com/time-series-models-streamhouse
zglg.work/en/ai/news/2026-11-18-ibm-s-time-series-foundation-models-now-run-natively-inside-confluent
Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky
Top comments (0)