NVIDIA Kumo Forecast Unifies Pretrained Time-Series Models with Historical Pattern Retrieval

NVIDIA has released Kumo Forecast, a multi-variate time-series forecasting library that fuses a pretrained Transformer backbone with a retrieval-augmented pathway. Formerly known as NV-Tesseract Forecasting, the model builds on the MOMENT-1-large encoder — roughly 340 million parameters kept frozen — while only the forecasting head is trained for the target task.

The library introduces Domain-Aware Representation and Retrieval (DARR), an optional mode that supplements direct model predictions with k-nearest-neighbor matches drawn from a separate historical context DataFrame. Developers can tune the blend weight between the direct forecast and the retrieved patterns, and the input window remains distinct from the retrieval corpus.

Why it matters for GPU / AI infrastructure

Kumo Forecast's architecture separates heavy representation learning (done once on GPU clusters) from lightweight task-specific fine-tuning and inference. This pattern — frozen foundation model plus trainable head plus optional CPU-friendly retrieval — aligns with the shift toward efficient, scalable serving on heterogeneous infrastructure. Teams running forecasting workloads on AiGpu cloud can pretrain once, then deploy many specialized heads with minimal compute.

An integrated interpretability framework produces horizon-specific explanations without modifying the forecaster, aiding compliance and trust in regulated sectors. The code and weights are available under Apache 2.0 at NVIDIA/Kumo-TS and nvidia/Kumo-Forecast respectively.