NVIDIA says Kumo Forecast uses the MOMENT-1-large backbone, an approximately 340-million-parameter encoder, plus a trainable forecasting head. Its optional DARR mode performs nearest-neighbor retrieval over a separate historical context_df and blends retrieved patterns with the direct forecast. The release is available under Apache 2.0, and the library also provides a DataFrame-first API and an interpretability bundle with horizon-specific lag and feature attributions.
The mechanism is a two-path forecast: recent observations enter the direct model path, while a separate archive supplies similar historical patterns for retrieval; configuration controls their contribution to the final prediction. This matters now because pretrained time-series representations alone cannot encode every local seasonal pattern, operating regime, or equipment history available to a deploying organization. The immediate implication is that teams can test whether retrieval improves forecast quality on their own held-out windows while retaining an explicit direct-forecast baseline.
The source describes architecture, interfaces, and recommended evaluation, but it does not provide independent benchmark results showing that DARR wins across domains or production evidence of reliability. Watch whether NVIDIA or users publish domain-specific comparisons between direct and retrieval-enhanced Kumo Forecast, including error, latency, and failure cases. This starts a distinct analytical record because no prior New Runtime record describing this Kumo Forecast release was supplied.