NVIDIA's Apache-2.0 Kumo Forecast adds nearest-neighbour retrieval to a frozen 340M-parameter MOMENT backbone, but publishes no accuracy numbers
Trained on 3 million data points, with per-horizon explanations and four benchmark datasets named. Verdict: worth a bake-off against your current forecaster, not a swap on faith.
Kumo Forecast is a multi-variate time-series forecasting model that combines a pretrained Transformer backbone with Domain-Aware Representation and Retrieval (DARR), a context-enhanced forecasting mode that blends direct model predictions with patterns retrieved from historical data.
Comprehensive take
An open, explainable forecaster with a retrieval option is a useful addition, and the Apache 2.0 licence lowers the cost of testing it. Without a single published error number, though, this is an invitation to benchmark, not evidence of improvement. Run it against your current model on your own data.
Use cases
Predictive maintenance
NVIDIA's Factory Operations Blueprint pairs Kumo Forecast with Kumo Anomaly in an early-access module, per the post.
Energy and load forecasting
The card evaluates on electricity-load and transformer-temperature datasets; check performance on your own series.
Explainable forecasts
Per-horizon lag and feature attributions can support review of why a forecast moved, without changing the model.
Specs
| Backbone | MOMENT-1-large Transformer encoder, ~340M parameters, frozen encoder and embedder; trainable forecasting head source |
|---|---|
| Licence | Apache 2.0 source |
| Training data | 3 million data points from the Monash Time Series Forecasting Archive (30 datasets) source |
| Evaluation datasets named | ECL, Traffic, ETTh1/ETTh2, ILI (no scores published) source |
| Input lengths | 256, 512, 1024 or 2048 observations; numeric data only source |
| Hardware and OS | Linux; NVIDIA Ampere and Hopper; tested on A100 and H100 source |
Insights
- Axion reading: the retrieval benefit is untested in public material; the post itself recommends comparing direct and retrieval modes on your own windows.
- Axion note: the model card's 04/07/2026 release date is ambiguous and precedes the October 21 blog, so the weights may have been available before the announcement.
What it is
Kumo Forecast, previously called NV-Tesseract Forecasting, is NVIDIA's open multivariate time-series forecaster. The base is the MOMENT-1-large Transformer encoder, about 340 million parameters, with encoder and embedder weights kept frozen and only a forecasting head trained. The addition is DARR, an optional mode that retrieves similar patterns from a separate historical DataFrame using k-nearest-neighbour search and blends them with the direct forecast. A model-agnostic interpretability layer returns lag and feature attributions per forecast horizon, plus stability diagnostics, without changing the forecaster. Code is at NVIDIA/Kumo-TS and weights at nvidia/Kumo-Forecast, both under Apache 2.0.
What the evidence does and does not show
The blog post and the model card give design detail but no results. The card lists four evaluation datasets (Electricity Load Diagrams, Traffic, ETTh1 and ETTh2, and the CDC influenza-like-illness series) and a training set of 3 million data points drawn from 30 datasets in the Monash Forecasting Archive, but neither document reports an error metric, a baseline, or a comparison with other open time-series models. The post itself closes by telling developers to compare the direct and retrieval-enhanced configurations on their own held-out windows. That is sound advice, and it is also an admission that the headline benefit of retrieval is not quantified here. NVIDIA authors also cite two IEEE papers on memory-augmented forecasting and a temporal-dependence error metric; we have not verified their results against this release.
One date detail: the model card gives a Hugging Face release date of 04/07/2026, while the blog is dated October 21, 2026. The format is ambiguous, so we cannot say whether the weights predate the announcement.
Practical limits
Input must be numeric time-series data with a timestamp column, at supported sequence lengths of 256, 512, 1024 or 2048 observations. In DARR mode the historical context must have dimensions similar to the input. The card lists Linux and NVIDIA Ampere or Hopper GPUs, and says tests ran on A100 and H100 systems. Horizons beyond the native length rely on autoregressive extension, which compounds error, so validate long horizons separately. The repository includes a fine-tuning example for a user-supplied CSV.
Who should try it
Teams doing predictive maintenance, energy-load or demand forecasting on sensor-style data, and who already have years of history to retrieve from, are the natural audience. NVIDIA points to its Factory Operations Blueprint, where early access to a predictive-maintenance module uses Kumo Anomaly and Kumo Forecast. If you have a working baseline, run the direct and DARR modes against it on held-out windows before changing anything in production.