Nvidia releases Kumo Forecast time-series model with retrieval and interpretability
Nvidia has open‑sourced Kumo Forecast, a multivariate time‑series forecasting model that blends a pretrained Transformer encoder with a retrieval‑augmented forecasting path. The model builds on the MOMENT‑1‑large backbone, using an approximately 340‑million‑parameter encoder that remains frozen while a trainable forecasting head learns the target task.
Key points
- Kumo Forecast uses a 340‑million‑parameter MOMENT‑1‑large encoder plus a trainable forecasting head.
- Domain‑Aware Representation and Retrieval (DARR) adds nearest‑neighbor pattern retrieval from a separate historical DataFrame.
- The model and weights are released under Apache 2.0 on Hugging Face, supporting up to 2,048 observation windows.
In Domain‑Aware Representation and Retrieval (DARR) mode, Kumo Forecast pulls similar patterns from a separate historical DataFrame via a k‑nearest‑neighbor search and merges those signals with the direct forecast. An interpretability framework can then generate horizon‑specific lag and feature attributions, semantic‑flow diagnostics, and trajectory‑stability analyses without changing the underlying forecaster. The library offers a DataFrame‑first API, configurable horizons, and support for input windows of 256, 512, 1,024 or 2,048 observations.
The model code (NVIDIA/Kumo-TS) and pretrained weights (nvidia/Kumo-Forecast) are hosted on Hugging Face under an Apache 2.0 license, allowing developers to run inference out‑of‑the‑box or fine‑tune on their own CSV data.
NVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability
Hugging Face · 21 October 2026
Loading the full article…
This text was published by Hugging Face. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
2sources- NVIDIA Announces Open Kumo Forecast Multivariate Time-Series ModelPress · Unite.AI ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- OpenAI says its AI agents weakened internal misalignment monitoring system · 10 src
- OpenAI agents removed safety monitors and denied actions, researcher says · 12 src
- Google launches Gemini 4 Argon in Japan with up to 1 million-token context · 2 src
- InternScience releases Sage-A1.5 35B agentic model for scientific tasks · 1 src
- Krafton releases Raon-1 mini preview, a 30B-A4B MoE model with 262,144 token context · 1 src
Comments
via GitHub Discussions