NVIDIA Releases Kumo Forecast, a Time-Series Model Combining Pretrained Representations with Historical Retrieval

·2026.10.21 21:51

Key point

The open-source model uses a MOMENT-1-large backbone and DARR mode to blend direct predictions with k-nearest-neighbor retrieval from historical data.

1 / 3

Details

  • NVIDIA Kumo Forecast (formerly NV-Tesseract Forecasting) is a multi-variate time-series forecasting model built on the MOMENT-1-large backbone with an approximately 340-million-parameter base encoder.
  • The model introduces Domain-Aware Representation and Retrieval (DARR), a mode that combines direct Transformer predictions with patterns retrieved from historical data using a k-nearest-neighbor approach.
  • A model-agnostic interpretability framework generates horizon-specific explanations, including lag and feature attributions, without modifying the underlying forecaster.
  • The library provides a DataFrame-first API, supports input sequence lengths of 256, 512, 1024, and 2048 observations, and is available under the Apache 2.0 license.
  • Code and weights are hosted on GitHub (NVIDIA/Kumo-TS) and Hugging Face (nvidia/Kumo-Forecast), with an example provided for fine-tuning on user-supplied CSV data.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.