Skip to content
Neural Daily

ModelsNews

NVIDIA Kumo Forecast: open time-series forecasting model

NVIDIA detailed Kumo Forecast, a time-series model that refines predictions with patterns retrieved from your own historical data and explains its forecasts. It ships under Apache 2.0.

By: Mikola Máté Patrik · · 2 min read

Industrial asset diagram: a pump linked to sensor time series, engineering documents, a 3D point cloud, camera images and a maintenance work order
Image: Cognite · Press kit · Official NVIDIA blog social image (og:image), slide marked Cognite 2026 · Source

What changed?

On October 21 NVIDIA published a detailed look (Hugging Face blog) at Kumo Forecast, previously known as NV-Tesseract Forecasting. The Kumo family was announced at GTC Berlin, and this is the first in-depth description of its forecasting member.

Kumo Forecast is a multivariate time-series forecasting model. It draws on two sources: representations learned in pretraining and the historical data of the application at hand.

How it works

  • Backbone: the MOMENT-1-large Transformer encoder, about 340 million parameters. The encoder stays frozen and only the forecasting head on top is trained.
  • DARR mode (Domain-Aware Representation and Retrieval): k-nearest-neighbor retrieval pulls similar patterns from a separately supplied historical DataFrame and blends them with the direct forecast. The weight of each path is configurable.
  • Interpretability: on request it returns lag and feature attributions per horizon, stability diagnostics and optional Integrated Gradients, without changing the forecaster.
  • Longer horizons: beyond its native range it extends forecasts autoregressively.

Limits to know

It only takes numeric time series: a timestamp column plus one or more value columns as a pandas DataFrame, CSV or JSON. Input length can be 256, 512, 1024 or 2048 observations. In DARR mode the context data must have a structure similar to the input.

According to the model card it runs on Linux with PyTorch and was tested on Ampere and Hopper GPUs. It was trained on about 3 million data points from the Monash Time Series Forecasting Archive. The blog gives no benchmark numbers, and NVIDIA asks for validation on your own data.

How to try it

Code is in the Kumo-TS GitHub repo and weights are on Hugging Face, both under Apache 2.0. The repo includes a fine-tuning example on your own CSV. NVIDIA also uses the model with Kumo Anomaly in an early access predictive maintenance module of its FOX factory blueprint.

Why it matters

Teams working on predictive maintenance or load planning get an open, commercially usable model that is not a black box. Because of retrieval, your own historical data can improve results without full retraining. Compare the direct and DARR modes on the same held-out windows before relying on either.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles