Nvidia's Kumo Forecast Tells You Why It Made Each Prediction, Not Just What It Predicts

Key points

• Nvidia released Kumo Forecast, a time-series forecasting model with roughly 340 million parameters, under a free Apache 2.0 licence.
• The model ships with a built-in interpretability framework that tells users which past data points drove each specific future prediction, an explanation most comparable tools do not produce.
• A retrieval mode called DARR lets the model search its own historical records for similar patterns before issuing a forecast, blending two sources of information into one output.
• Nvidia demonstrated the model in a factory predictive-maintenance setting, where it works alongside a companion anomaly-detection model called Kumo Anomaly.
• Model weights are publicly available on Hugging Face at nvidia/Kumo-Forecast under the project name NVIDIA/Kumo-TS.

Time-series forecasting, the task of predicting future numbers from a sequence of past measurements, sits behind some of the most consequential decisions in industry: when a machine will fail, how much stock to order, whether a power grid will hit capacity. Until now, the AI models doing this work have mostly been black boxes. You get a number. You don't get a reason.

Nvidia's new model, Kumo Forecast (previously called NV-Tesseract Forecasting), is designed to change that.

How does it actually work?

Kumo Forecast combines four parts. A pretrained encoder called MOMENT-1-large, a Transformer architecture (the same family of neural-network design that powers most modern AI assistants), reads the incoming data. A separate forecasting head, trained on top of that frozen encoder, translates the AI's internal understanding into future values.

For most tasks that is enough. But Nvidia added an optional second pathway called Domain-Aware Representation and Retrieval, or DARR. When switched on, DARR searches a bank of historical data the developer supplies and finds the patterns closest to what the model is currently seeing, a technique called nearest-neighbour retrieval. It then blends those retrieved examples with the direct forecast. Think of it as the model checking its own memory before committing to an answer.

The research behind DARR is published at arxiv.org. A companion paper on the interpretability framework, NVExplain, is cited in the model documentation at IEEE Xplore.

Why does the explainability piece matter?

It matters because regulators and engineers alike increasingly want a justification, not just a forecast.

Kumo Forecast's interpretability layer runs alongside the model without changing the underlying forecaster. It produces what Nvidia calls an "explanation bundle": a matrix showing which historical lag, meaning which past data point at which distance back in time, most influenced each future prediction horizon. A factory engineer can see, in plain terms, which time window drove today's anomaly flag rather than simply accepting a probability score.

This is genuinely new. Explainability has lagged badly behind raw forecasting performance across the industry, a pattern we tracked in our 15 October report on Mistral's Timestral, where benchmark performance again led and transparency trailed. Nvidia's approach is model-agnostic, meaning it can wrap around other forecasters too, not just Kumo.

Nvidia showed the model working inside its Factory Operations Blueprint, a pre-built industrial AI template, where Kumo Forecast pairs with a separate anomaly-detection model to flag equipment failures before they happen.

What are the limits?

The model only accepts numeric, time-stamped data. Supported sequence lengths run from 256 to 2,048 observations. DARR mode requires the historical context data to share the same structure as the input. Nvidia explicitly warns that developers must validate the model on their own use-case data before deploying it.

Fine-tuning on custom data is supported and documented in the NVIDIA/Kumo-TS repository. The Apache 2.0 licence means commercial use is permitted without royalties.

For industries where a wrong forecast has real consequences, the explainability layer is the practical story here. A number you can't defend is hard to act on. Kumo Forecast at least lets you show your work.

Common questions

Do I need Nvidia hardware to run this?

The model card and documentation do not restrict the model to Nvidia GPUs. Because the weights are open and the licence is Apache 2.0, developers can run inference on any compatible hardware, though Nvidia's own chips will typically give the fastest results.

What kind of data can I feed it?

Kumo Forecast requires numeric data with a timestamp column. It accepts a pandas DataFrame (a standard Python data table) or a CSV or JSON file. Non-numeric inputs, including text and images, are not supported in the current release.