NVIDIA open-sources Kumo Forecast time-series model under Apache 2.0

NVIDIA Kumo Forecast (formerly NV-Tesseract Forecasting) is now available as open source under Apache 2.0. The model combines a frozen MOMENT-1-large Transformer encoder with a trainable forecasting head, and offers an optional retrieval-augmented mode called DARR that blends direct predictions with historical pattern matching. It ships with interpretability tools and a pandas-native API, with model weights and code hosted on Hugging Face.

Architecture and retrieval mode

• Backbone: MOMENT-1-large Transformer encoder (~340M parameters), kept frozen during fine-tuning
• Forecasting head: Trainable layer maps representations to future values; supports autoregressive extension for longer horizons
• DARR mode: k-nearest-neighbor retrieval from a separate historical context DataFrame, blended with direct forecast via configurable weights
• Input constraints: Numeric temporal data only; sequence lengths of 256, 512, 1024, or 2048 observations
• Interpretability: Horizon-specific lag/feature attributions, semantic-flow and trajectory-stability diagnostics, optional embedding Integrated Gradients
• Distribution: Model code at NVIDIA/Kumo-TS, weights at nvidia/Kumo-Forecast on Hugging Face

Why it matters

The release adds a commercially permissive, pretrained time-series foundation model to the open AI ecosystem, with retrieval augmentation and built-in interpretability features that are uncommon in existing open forecasting tools.