NVIDIA Releases Kumo Forecast

NVIDIA has released Kumo Forecast, an open-source forecasting library for multivariate time-series data, under the Apache 2.0 license. The model, formerly called NV-Tesseract Forecasting, combines a pretrained Transformer with an optional retrieval system that adds related patterns from historical data. Its base encoder has approximately 340 million parameters and uses the MOMENT-1-large backbone; developers can access the code as NVIDIA/Kumo-TS and the weights as nvidia/Kumo-Forecast.

In its direct forecasting mode, Kumo Forecast uses recent observations and a trained forecasting head to predict future values. In Domain-Aware Representation and Retrieval (DARR) mode, it also finds similar patterns in historical data using a nearest-neighbor method, then combines them with the direct prediction. Developers provide recent observations in the input window and can supply a separate historical DataFrame for retrieval. Configuration options control how much the final prediction relies on each pathway.

The model keeps its pretrained encoder and embedder weights frozen while training the forecasting head. For forecasts longer than the model’s native horizon, it can extend predictions autoregressively, feeding earlier predictions into later steps. The library offers a DataFrame-first interface, configurable forecast horizons and optional historical context.

An optional interpretability framework works alongside the forecaster and does not require a different forecasting architecture. It produces explanations for each forecast horizon, including lag and feature attributions, semantic-flow and trajectory-stability diagnostics, and optional embedding Integrated Gradients. NVIDIA links the DARR approach to the published paper “Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains” and the explanation framework to “NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates.”

Kumo Forecast accepts numeric time-series data with a timestamp column and one or more value columns, supplied as a pandas DataFrame or CSV/JSON. Supported input sequence lengths are 256, 512, 1024 and 2048 observations. For DARR, the historical context must have dimensions similar to the input context. NVIDIA recommends comparing direct and retrieval-enhanced configurations on the same held-out forecasting windows and validating quality across intended horizons and operating conditions. The Kumo-TS repository also includes an example for fine-tuning the model on a user-supplied CSV. A predictive-maintenance demo in the Factory Operations (FOX) Blueprint uses Kumo Anomaly and Kumo Forecast to analyze real-time data and predict future anomalies.

NVIDIA has released Kumo Forecast, an open-source forecasting library for multivariate time-series data, under the Apache 2.0 license. The model, formerly called NV-Tesseract Forecasting, combines a pretrained Transformer with an optional retrieval system that adds related patterns from historical data. Its base encoder has approximately 340 million parameters and uses the MOMENT-1-large backbone; developers can access the code as NVIDIA/Kumo-TS and the weights as nvidia/Kumo-Forecast.

In its direct forecasting mode, Kumo Forecast uses recent observations and a trained forecasting head to predict future values. In Domain-Aware Representation and Retrieval (DARR) mode, it also finds similar patterns in historical data using a nearest-neighbor method, then combines them with the direct prediction. Developers provide recent observations in the input window and can supply a separate historical DataFrame for retrieval. Configuration options control how much the final prediction relies on each pathway.

The model keeps its pretrained encoder and embedder weights frozen while training the forecasting head. For forecasts longer than the model’s native horizon, it can extend predictions autoregressively, feeding earlier predictions into later steps. The library offers a DataFrame-first interface, configurable forecast horizons and optional historical context.

An optional interpretability framework works alongside the forecaster and does not require a different forecasting architecture. It produces explanations for each forecast horizon, including lag and feature attributions, semantic-flow and trajectory-stability diagnostics, and optional embedding Integrated Gradients. NVIDIA links the DARR approach to the published paper “Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains” and the explanation framework to “NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates.”

Kumo Forecast accepts numeric time-series data with a timestamp column and one or more value columns, supplied as a pandas DataFrame or CSV/JSON. Supported input sequence lengths are 256, 512, 1024 and 2048 observations. For DARR, the historical context must have dimensions similar to the input context. NVIDIA recommends comparing direct and retrieval-enhanced configurations on the same held-out forecasting windows and validating quality across intended horizons and operating conditions. The Kumo-TS repository also includes an example for fine-tuning the model on a user-supplied CSV. A predictive-maintenance demo in the Factory Operations (FOX) Blueprint uses Kumo Anomaly and Kumo Forecast to analyze real-time data and predict future anomalies.