NVIDIA Open-Sources Kumo Forecast, Pairing Pretrained Time-Series Representations With Pattern Retrieval

NVIDIA published Kumo Forecast, previously known as NV-Tesseract Forecasting, in a Hugging Face blog post on October 21. The model targets multivariate time-series forecasting, and its premise is that a forecaster should draw on two sources at once: representations learned during pretraining, and the historical observations a given business already holds.

The backbone is the MOMENT-1-large Transformer encoder, with an approximately 340-million-parameter base encoder. Encoder and embedder weights stay frozen while only the forecasting head is trained, keeping the pretrained representation model separate from the task-specific prediction layer.

Beyond the direct path, Kumo Forecast offers Domain-Aware Representation and Retrieval, or DARR, which runs a nearest-neighbor search over historical context and blends the retrieved patterns with the model's own prediction. The input window and the retrieval context are separate inputs, and developers can configure how strongly the final forecast leans on each pathway.

The library ships a DataFrame-first API and configurable forecast horizons; horizons beyond the model's native range can be extended autoregressively. Retrieval context arrives through its own context_df argument, which keeps recent observations and historical reference data distinct at the data layer.

Interpretability is the third pillar. The framework is model-agnostic and produces horizon-specific explanations covering lag and feature attributions without modifying the underlying forecaster, so the output is an explanation bundle rather than a rebuilt model.

NVIDIA names predictive maintenance and platform integration as demonstration scenarios, and the post includes a dedicated limitations section while citing the paper “Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains” for the DARR approach. That framing points at the trade-off of retrieval augmentation: the richer the history, the more the answer depends on getting the match and the context governance right.

Kumo-Forecast is released under Apache 2.0, with model code at NVIDIA/Kumo-TS and weights at nvidia/Kumo-Forecast. For teams already running MOMENT-family baselines, that makes it a low-friction option rather than a rebuild.

The more consequential choice is packaging. NVIDIA shipped retrieval and explainability as library features rather than a bare checkpoint, which pushes time-series forecasting toward an operable system component. What to watch is how much DARR adds over plain fine-tuning on messy production data, and how retrieval cost scales as the context grows.