NVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability

What it claims

Kumo Forecast (formerly NV-Tesseract Forecasting) is NVIDIA's multivariate forecaster. It pairs a frozen MOMENT-1-large encoder (about 340M parameters) with a trained forecasting head and an optional retrieval mode called DARR. The page gives no benchmark numbers, baselines or compute figures. Instead, it points to two papers for method details: "Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains" for DARR, and "NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates" for interpretability.

How it works

The encoder and embedder weights stay frozen. Only the forecasting head is trained. The head predicts the requested horizon directly. Horizons beyond the model's native range use autoregressive extension, so long forecasts are recursive rather than generated in one pass.

DARR keeps two inputs separate. The input window supplies recent observations. A separate historical DataFrame, passed as context_df, is searched with k-nearest-neighbor retrieval. The retrieved patterns are blended with the direct forecast, and a configuration setting controls how much weight each pathway gets.

flowchart LR
  W[Input window] --> E[Frozen MOMENT encoder]
  E --> H[Trained head]
  C[context_df history] --> K[kNN retrieval]
  H --> B{Blend weight}
  K --> B
  B --> F[Forecast]
  F -.optional.-> X[Explanation bundle]

The interpretability framework is model-agnostic and does not change the forecaster. When enabled, it returns an explanation bundle with per-horizon lag and feature attributions, semantic-flow diagnostics, trajectory-stability diagnostics, and optional embedding Integrated Gradients.

How to use it

Code is in the GitHub repository NVIDIA/Kumo-TS and weights are at nvidia/Kumo-Forecast on Hugging Face, both under Apache 2.0. The API is DataFrame-first. Input is a timestamp column plus one or more numeric value columns, supplied as a pandas DataFrame or as CSV or JSON. The repository includes an example of fine-tuning on a user-supplied CSV. The page gives no install or run commands.

NVIDIA's suggested evaluation is to run the direct and DARR configurations on the same held-out windows and check quality across the intended horizons. It also says use-case-specific testing is required. The Factory Operations (FOX) Blueprint uses Kumo Forecast with Kumo Anomaly in an Early Access Predictive Maintenance module. A GTC San Jose 2026 session shows platform integration.

Limits

Input length must be exactly 256, 512, 1024 or 2048 observations. These are discrete sizes, not a continuous range. Only numeric temporal data is supported. In DARR mode, the context data must have dimensions similar to the input. Forecasts past the native horizon are autoregressive, and the page does not quantify how error grows with that extension. The page reports no accuracy results, so performance claims rest on the linked papers.