NVIDIA Kumo Forecast Adds Historical Retrieval to Time-Series AI
NVIDIA opens Kumo Forecast, a 340-million-parameter time-series model blending MOMENT predictions, historical retrieval and horizon-level explanations.
Summary
NVIDIA released Kumo Forecast, formerly NV-Tesseract Forecasting, on October 21, 2026. The multivariate time-series model combines MOMENT-1-large, an approximately 340-million-parameter Transformer encoder, with a trainable forecasting head. Encoder and embedder weights remain frozen. Direct mode predicts from recent inputs, while Domain-Aware Representation and Retrieval, or DARR, uses k-nearest-neighbor retrieval from a separate historical context_df. Developers can tune how strongly direct and retrieved patterns influence predictions.
The DataFrame-first library supports configurable horizons, autoregressive extension, numeric temporal data in pandas DataFrames, CSV or JSON, and input lengths of 256, 512, 1024 or 2048 observations. Its model-agnostic explanations provide horizon-specific lag and feature attributions, semantic-flow and trajectory-stability diagnostics, and optional embedding Integrated Gradients. DARR requires context dimensions similar to the input. Apache 2.0 code is available at NVIDIA/Kumo-TS, weights at nvidia/Kumo-Forecast, and the repository includes CSV fine-tuning guidance. The Factory Operations, or FOX, Blueprint’s early-access Predictive Maintenance module pairs Kumo Forecast with Kumo Anomaly to predict future anomalies from real-time data. NVIDIA recommends comparing direct and DARR configurations on held-out windows across intended horizons and operating conditions. Supporting research includes Memory-Augmented Forecasting and NVExplain.
Positives
- Approximately 340 million parameters in MOMENT-1-large provide pretrained temporal representations while a separate forecasting head handles task-specific prediction.
- DARR combines direct forecasts with k-nearest-neighbor retrieval from historical data, with configurable weighting between both pathways.
- Horizon-specific explanations include lag and feature attributions, semantic-flow diagnostics, trajectory-stability diagnostics and optional embedding Integrated Gradients.
- Apache 2.0 licensing covers Kumo Forecast, with code at NVIDIA/Kumo-TS and weights at nvidia/Kumo-Forecast.
- FOX Blueprint’s early-access Predictive Maintenance module combines Kumo Forecast and Kumo Anomaly to predict future anomalies from real-time data.
- CSV fine-tuning guidance gives developers a documented route from pretrained inference to adaptation on proprietary data.
Risks & concerns
- Numeric temporal data is the only supported input, excluding categorical, textual and other unstructured information.
- 256, 512, 1024 and 2048 observations are the only supported input sequence lengths.
- DARR requires historical context dimensions similar to those of the forecast input.
- Horizons beyond the model’s native capability depend on autoregressive extension, which recursively extends earlier predictions.
- Use-case-specific testing remains necessary across held-out windows, intended horizons and operating conditions.