NVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability

NVIDIA released Kumo Forecast, a multivariate time-series forecasting model built on the MOMENT-1-large backbone with an approximately 340-million-parameter frozen encoder and a trainable forecasting head. Its DARR mode blends direct forecasts with k-nearest-neighbor retrieval from historical context, and an interpretability framework adds per-horizon lag and feature attributions. The model is available under Apache 2.0, with code at NVIDIA/Kumo-TS and weights at nvidia/Kumo-Forecast.