Time-series & structured AI

NVIDIA opens Kumo Forecast with retrieval-enhanced time-series AI

Kumo Forecast pairs pretrained temporal representations with historical-neighbor retrieval and separate interpretability outputs.

Open Model Weights published22 Oct 2026
Primary sourceNVIDIA official Kumo Forecast engineering announcement
Source published2026-10-21
FOLLOW OUR COVERAGE

Follow Open Model Weights on GoogleSee more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original 1200 by 675 editorial technical schematic: RECENT MEASUREMENTS through MOMENT + FORECAST HEAD through HORIZON FORECAST, detailing 340M-CLASS FROZEN ENCODER; DARR · HISTORICAL NEIGHBORS; LAG + FEATURE ATTRIBUTIONS.
Original Open Model Weights vector explainer of the publisher-described workflow; conceptual only, not a documentary photo or an independent speed or accuracy benchmark. AI-generated editorial illustration; not a documentary photograph.

NVIDIA introduced Kumo Forecast on October 21 as an open time-series forecasting model with an optional historical-retrieval pathway. The release goes beyond ordinary language generation: it targets multivariate measurements such as demand, telemetry or other sequential business signals. Its most distinctive design choice is to combine a pretrained temporal representation model with patterns drawn from a user's own historical dataset instead of relying only on what was learned during pretraining.

The released architecture builds on the approximately 340-million-parameter MOMENT-1-large encoder and adds a trainable forecasting head. NVIDIA describes freezing the underlying encoder and embedder during head training. The direct pathway turns recent observations into a forecast for a configured horizon. When applications need forecasts beyond the base horizon, the SDK can extend them autoregressively; these recursive predictions can accumulate error and deserve separate tests.

An optional Domain-Aware Representation and Retrieval, or DARR, pathway uses nearest-neighbor search over a separate historical DataFrame. This retrieves earlier patterns and blends their signal with the direct model prediction. An associated interpretability layer supplies horizon-specific feature and lag attributions, trajectory diagnostics and optional embedding Integrated Gradients. These tools help investigate a prediction, but an attribution output should not automatically be treated as proof of causality.

NVIDIA identifies Kumo Forecast as Apache-2.0 and provides code and weight links. Our assessment: retrieval-augmented forecasting is a relevant open-weight option for organizations with rich local history, but benefits will vary with seasonality, missing values, data leakage, distribution shift and how similar the retrieved windows really are. Benchmarking against simple seasonal and statistical baselines remains essential, and no independent Open Model Weights accuracy result is claimed.

OMW REGISTRY WATCH

What this changes for the evidence layer

Verify exact nvidia/Kumo-Forecast checkpoint revision, Apache-2.0 licensing, MOMENT backbone provenance, DARR context licensing and the tested horizon before canonical registry intake.

PRIMARY SOURCE

NVIDIA official Kumo Forecast engineering announcement

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗