NVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability
huggingface ·
NVIDIA has introduced Kumo Forecast, a multivariate time-series forecasting model formerly known as NV-Tesseract Forecasting. Built on the MOMENT-1-large Transformer backbone with approximately 340 million parameters, the model combines pretrained representations with a trainable forecasting head. It features a unique Domain-Aware Representation and Retrieval (DARR) mode that blends direct predictions with patterns retrieved from historical data using a nearest-neighbor approach. The library offers a DataFrame-first API and supports configurable forecast horizons, including autoregressive extension for longer predictions. A key innovation is its model-agnostic interpretability framework, which provides horizon-specific explanations such as lag attributions and semantic-flow diagnostics without altering the core architecture. Kumo Forecast is available under the Apache 2.0 license and includes demos for predictive maintenance and platform integration. Developers can access the model code and weights via Hugging Face to start building or fine-tuning the system on their own datasets.
huggingface