NVIDIA releases open-source multivariate time series forecasting model Kumo Forecast

On October 21, 2026, NVIDIA released the open-source multivariate time series forecasting model Kumo Forecast. The model uses the Apache 2.0 license, with code in the NVIDIA/Kumo-TS repository and pretrained weights hosted on Hugging Face at nvidia/Kumo-Forecast. The model uses a MOMENT-1-large Transformer encoder with approximately 340 million parameters as a frozen backbone, training only the forecasting head, supports input sequence lengths from 256 to 2048, and offers an optional DARR retrieval-augmented mode. Its interpretability framework can output results such as lag and feature attributions and semantic flow diagnostics. Running it requires Python 3.10 or later and PyTorch 2.7 or later, and the official recommendation is to compare the direct forecasting and retrieval-augmented configurations on the same held-out window. There are currently no subsequent development facts.

AI-generated from public reports

1. Oct 21· 1 reports

   NVIDIA releases open-source multivariate time series forecasting model Kumo Forecast

   Source:Unite.AI