Nvidia released Kumo Forecast, an open-source multivariate time-series forecasting model that combines a pretrained Transformer with historical pattern retrieval, according to technical documentation published by Hugging Face.
The system, formerly known as NV-Tesseract Forecasting, is built on the MOMENT-1-large backbone under an Apache 2.0 license. It pairs an approximately 340-million-parameter base encoder with a trainable forecasting head. During training, the architecture keeps encoder and embedder weights frozen, separating the pretrained representation model from the task-specific prediction layer. Nvidia published the model code in the NVIDIA/Kumo-TS repository and hosted the weights under nvidia/Kumo-Forecast.
The model provides two forecast mechanisms. In direct mode, the model maps representations to future values for requested horizons, using autoregressive extension to generate forecasts that exceed its native horizon limits. In Domain-Aware Representation and Retrieval (DARR) mode, the system retrieves patterns from historical context through a nearest-neighbor search and blends them into the direct prediction. The library accepts historical context through a separate context_df argument, requiring the context dimensions to match the input.
Inputs must contain temporal, numeric data structured with a timestamp column and one or more numeric columns in pandas DataFrame, CSV, or JSON format. Supported sequence lengths include 256, 512, 1024, and 2048 observations. Alongside forecasts, the library includes a model-agnostic interpretability framework. This tool generates horizon-specific explanation bundles covering lag attributions, feature attributions, semantic-flow diagnostics, trajectory-stability diagnostics, and optional embedding Integrated Gradients.
Nvidia demonstrated the model in its Factory Operations Blueprint, pairing Kumo Forecast with Kumo Anomaly to predict future operational issues in an early-access predictive maintenance module. The company also supplied fine-tuning workflows in the Kumo-TS repository for user-supplied CSV datasets, while requiring application-specific testing across direct and retrieval-enhanced modes on held-out windows.
