The Pulse
NVIDIA Details Kumo Forecast’s Retrieval-Based Time-Series Model
NVIDIA’s October 21 technical post explains how Kumo Forecast combines a pretrained time-series model with retrieval of similar historical patterns. The Apache 2.0 model supports numeric time-series inputs and includes tools for explaining
AI.info Team · Written with AI · fact-checked ·
“In Domain Aware Representation and Retrieval (DARR) mode, it also retrieves related patterns from historical data using a nearest-neighbor approach and combines that information with the direct forecast.”
Saira Qureshi, Aditi Gautam, Manikandan Ravikiran, Sena Ekiz and Viraj P. Modak, NVIDIA
NVIDIA’s October 21 technical post describes Kumo Forecast, a multivariate time-series model that pairs a pretrained Transformer with a retrieval pathway for historical patterns. The post presents the model as a way to combine what a model learned during pretraining with relevant examples from a developer’s own time-series data. It is a technical explanation, not the model’s initial release: NVIDIA’s Hugging Face model card lists April 7, 2026, as its release date.
Two paths to a forecast
Kumo Forecast’s direct path uses a Transformer based on MOMENT-1-large and a trainable forecasting head. NVIDIA says the encoder and embedder weights stay frozen while the forecasting head is trained, separating the pretrained representation from the task-specific prediction layer. The model has an approximately 340-million-parameter base encoder, plus that forecasting head.
DARR, short for Domain-Aware Representation and Retrieval, adds a second source of information. It uses a nearest-neighbor approach to retrieve patterns from historical context data, then combines those patterns with the direct prediction. The input window and retrieval context are distinct: recent observations go into the forecast, while a separate historical DataFrame can provide material for retrieval.
Developers can configure how much the final forecast relies on each path. The design means retrieval is an option rather than a mandatory stage of every prediction. NVIDIA’s post recommends comparing the direct and retrieval-enhanced configurations on the same held-out forecasting windows, then validating results against the relevant forecast horizons and operating conditions.
Interpretability comes alongside the prediction
The library also offers a model-agnostic interpretability framework. When enabled, it can return lag and feature attributions for each forecast horizon, semantic-flow and trajectory-stability diagnostics, and optional embedding Integrated Gradients. NVIDIA describes these outputs as accompanying the forecast without changing the forecasting architecture.
That distinction matters for deployment: explanation outputs may help teams inspect what influenced a prediction, but they do not establish that a forecast is accurate for a particular operation. The post directs developers to evaluate performance on their own use case rather than treating the model description as evidence of a universal accuracy advantage.
Numeric data, defined input lengths
Kumo Forecast accepts numeric temporal data with a timestamp column and one or more numeric value columns. Supported input sequences are 256, 512, 1,024 or 2,048 observations, supplied as a pandas DataFrame or as CSV or JSON. In DARR mode, NVIDIA says the dimensions of the historical context should be similar to those of the input context.
For forecast horizons beyond the model’s native capability, the library supports autoregressive extension, in which predictions are extended recursively. That is distinct from a forecast generated directly for the requested horizon. The model card also cautions that deployment requires additional testing with use-case-specific data.
Code and weights are available under Apache 2.0
NVIDIA makes the Kumo-TS code and Kumo Forecast weights available through Hugging Face, and the model is licensed under Apache 2.0. The accompanying library uses a DataFrame-first API and includes an example for fine-tuning on a user-provided CSV. The model card describes a forecasting system for numeric time series; it does not, by itself, establish that retrieval improves results across datasets.
For teams considering the model, the practical test is therefore a controlled comparison: run direct and DARR-enabled forecasts on the same held-out periods, examine the intended horizons, and assess the outputs against local operating requirements.