NVIDIA releases open Kumo Forecast time-series model

NVIDIA announced Kumo Forecast, an open multivariate time-series forecasting model built on the 340-million-parameter MOMENT-1-large Transformer backbone. Code and weights are available under Apache 2.0 on GitHub and Hugging Face, with optional retrieval over historical data and a model-agnostic interpretability framework.

• Built on MOMENT-1-large with a ~340M-parameter encoder
• Code and weights released under Apache 2.0 on GitHub and Hugging Face
• Supports horizons of 256–2048 observations and DARR kNN retrieval mode
• Runs on PyTorch with Ampere and Hopper GPUs, tested on A100 and H100