Realtime AI News
IBM's time-series foundation models now run natively inside Confluent
IBM Research and Confluent have put IBM's Granite time-series foundation models directly inside Confluent, callable as built-in SQL functions in the cloud or on premises. IBM says the model family has passed 50 million downloads and runs on CPUs, pushing forecasts and anomaly detection into the stream itself.

IBM Research used a Hugging Face blog post on November 18 to detail an integration with Confluent that puts IBM's time-series foundation models natively inside the data streaming platform. In Confluent Cloud the models are fully managed; on premises they run through Confluent Platform for Apache Flink. Instead of standing up a separate model service, teams invoke forecasting and anomaly detection from SQL against data already moving through the pipeline.
The two documented entry points are AI_FORECAST and AI_DETECT_ANOMALIES. The first returns a forecast with a range for a given series; the second builds an expected range from recent history and flags new readings that fall outside it at a confidence level the team chooses. Both draw on pretrained models, including TTM, PatchTST-FM and TSPulse, so a separate model does not have to be trained and maintained for every series.
Scale and footprint are the pitch. IBM says the four-model family has exceeded 50 million downloads, and that the models are small enough to run on standard CPUs alongside the streaming pipeline rather than on dedicated GPU infrastructure. For organizations that already run Kafka and Flink, that is a meaningfully lower barrier to adding a predictive layer than provisioning a separate inference stack.
The post walks through concrete scenarios. On an ice cream line, a TTM model reads SCADA telemetry and QA measurements every ten minutes and forecasts the share of tubs expected to be in spec four hours ahead; when the forecast drifts below target, an optimization step scores combinations of freezer draw temperature and airflow for the operator to approve. In retail, each sales series gets its own expected range instead of a fixed threshold that fires on predictable weekend peaks.
The supply chain and financial examples lean toward agents. A demand forecast can return a 28-day, P10-P90 range refreshed as sales arrive, feed a reorder point calculation, and have an agent draft a purchase order for human review. For fraud, every card is scored against its own spending history and out-of-range transactions are passed to the bank's existing systems as an extra signal. On a trading desk, PatchTST-FM forecasts 30 minutes of volume on each one-minute bar to pace a large order.
IBM frames all of this inside Confluent's Streamhouse architecture: connect, stream, process, govern and query capabilities keep current business context available to applications and agents, Flink transforms and enriches data in flight, Lightning Tables serve low-latency current state, and the time-series models sit in an intelligence layer alongside large language models and statistical methods.
One caution is worth stating. The blog describes what organizations could build, which reads closer to a reference architecture than to named production customers, and the model code is public in the ibm-granite/granite-tsfm repository. The things to watch next are pricing and quotas for the built-in functions in Confluent Cloud, forecast quality on real enterprise data, and whether the draft-then-approve agent loop holds up outside simulated examples.
Why it matters
The move shifts time-series foundation models from a model-centric release to a native platform capability, lowering the cost of real-time forecasting for data teams and moving competition toward data quality, pipeline configuration and human-in-the-loop approval design.
Nearby Updates
All11/18, 23:01
signageOS taps Anthropic's Claude Mythos to secure 180,000 screens
Digital signage device-management company signageOS has joined Anthropic's Cyber Verification Program, gaining access to the Claude Mythos model for defensive security work. The company is applying it to the cloud services, APIs and software behind more than 180,000 connected devices across over 50 hardware platforms.
11/18, 22:34
Vero Security raises $11M to govern AI agents at runtime
Vero Security, a New York-based cybersecurity startup, said on November 18 that it raised $11 million led by TLV Partners, with GTM Capital, Maple Capital and angels participating. Founded in 2025, the company is building an AI Runtime Governance platform that continually reassesses what agents can access and do, then proposes fixes for security teams to approve.
11/18, 23:48
Altman Says He Doesn't Know if Anthropic's AI Development Is Safe, a Day Before Its IPO
At The New York Times' DealBook Summit, OpenAI CEO Sam Altman said he does not know whether Anthropic's approach to building AI is safe, calling a race run in the fog of war a bad thing. The remark landed a day before Anthropic's expected market debut, alongside calls for shared safety standards and more transparency.
11/18, 22:00
NVIDIA RTX Spark compact desktops go on sale with up to 128GB of unified memory
NVIDIA's RTX Spark compact desktops are now on sale, with ASUS, Dell and MSI shipping systems that carry up to 128GB of unified memory. The machines target local AI agents, image and video generation and model fine-tuning, turning a build-it-yourself workflow into an off-the-shelf Windows PC.