Kumo Forecast pairs a frozen encoder with lookups of a company’s own history

Nvidia has released Kumo Forecast, a multivariate time-series model previously called NV-Tesseract Forecasting, under an Apache 2.0 license with code on GitHub and weights on Hugging Face. It sits on the MOMENT-1-large backbone, an existing open family of time-series foundation models; Nvidia keeps the roughly 340-million-parameter encoder frozen and trains only a forecasting head. The repository includes an example for fine-tuning on a user-supplied CSV.

The distinguishing feature is a mode Nvidia calls Domain-Aware Representation and Retrieval. Alongside its direct prediction, the model can pull similar stretches from a separate historical dataset using nearest-neighbor search and blend the two, with operators choosing how much the final number leans on each path. It also produces per-horizon explanations showing which lags and features drove the forecast, plus stability diagnostics. Nvidia links the retrieval mode and the explanation framework to two separate research papers. The release accompanies Kumo Relational, a sibling model for relational databases (Kumo Relational).

Forecasting underpins demand, energy load and equipment-wear planning. Pretrained time-series models promise to generalize across domains, but every plant, grid and warehouse has its own rhythms. Letting the model consult a company’s own history at prediction time acknowledges that general representations only go so far, and resembles how an experienced engineer forecasts: what does this look like, and what happened the last time it looked like this?

The explanation layer may matter most in practice. Nvidia says the explanations run alongside the forecast without changing the forecaster. In regulated or safety-sensitive settings, showing which inputs mattered is often what separates a pilot from a deployment. Nvidia has pursued similar structured approaches to agent memory (NemoClaw memory agent).

Nvidia lists the limits plainly. Inputs must be numeric, input windows come in fixed sizes from 256 to 2,048 observations, and the retrieval context must resemble the input’s shape. Longer horizons are reached by extending predictions recursively, and recursive forecasts tend to drift. The post offers no head-to-head accuracy figures, so this is a capability announcement rather than a benchmark result, and Nvidia advises adopters to test direct and retrieval modes on their own held-out windows.

Retrieval can also make a forecast look grounded because it points to real past episodes, but history misleads when conditions shift, and a well-explained wrong forecast can be more persuasive than an unexplained one. Explanations are evidence to examine, not a substitute for accuracy. Strategically, Kumo Forecast already appears in an early-access predictive maintenance module alongside a sibling anomaly model, fitting Nvidia’s pattern of seeding open models inside its industrial blueprints. More in our Nvidia coverage, AI tools coverage and enterprise coverage.