NVIDIA releases Kumo Forecast time series prediction model
The story so far
NVIDIA has released a multi-variable time series forecasting model called Kumo Forecast (formerly NV-Tesseract Forecasting) and has made the model weights public, using the Apache 2.0 license. The model uses MOMENT-1-large as its core, with approximately 340 million parameters in the encoder plus a trainable prediction head, with encoder and embedding layer weights frozen. According to NVIDIA, the model combines a pre-trained Transformer core with the 'Domain-Aware Representation and Retrieval' (DARR) mode: DARR mixes the model's direct predictions with patterns retrieved from historical data to enhance predictions. When using the DARR mode, the context dimension needs to be close to the input context; supported input sequence lengths are 256, 512, 1024, and 2048 observations.
AI-written from the reports · Updated 28 days ago
Coverage timeline
Follow the coverage to see every side of the story.
- Hugging Face BlogNVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability
NVIDIA released the multivariate time-series forecasting model Kumo Forecast (formerly NV-Tesseract Forecasting), featuring a backbone of MOMENT-1-large with approximately 340 million parameters in the encoder plus a trainable forecasting head, with encoder and embedding layer weights frozen.