NVIDIA releases Kumo Forecast, an Apache 2.0 time-series model built on a ~340M-parameter MOMENT-1-large encoder

NVIDIA has released Kumo Forecast, a multivariate time-series forecasting model that combines a pretrained Transformer backbone with k-nearest-neighbor retrieval over historical data and per-horizon interpretability outputs. The model, formerly known as NV-Tesseract Forecasting, is available under the Apache 2.0 license. NVIDIA published the announcement on Hugging Face on October 21, 2026.

Specifications

• Backbone: MOMENT-1-large Transformer encoder, approximately 340 million parameters in the base encoder, plus a trainable forecasting head
• Training approach: Encoder and embedder weights stay frozen while the forecasting head is trained
• Supported input lengths: 256, 512, 1024, and 2048 observations
• Input format: Timestamp column plus one or more numeric value columns, supplied as a pandas DataFrame or CSV/JSON
• License: Apache 2.0
• Code: NVIDIA/Kumo-TS; weights: nvidia/Kumo-Forecast
• Pricing: Not applicable for the open weights; no hosted API pricing disclosed
• Benchmark scores: Not disclosed in the announcement
• Training data and cutoff: Not disclosed in the announcement

Two forecasting paths

In direct mode, the MOMENT-based encoder feeds a forecasting head that predicts future values for the requested horizon. For horizons beyond the model's native capability, the library supports autoregressive extension, which extends predictions recursively.

In Domain-Aware Representation and Retrieval (DARR) mode, the model also retrieves related patterns from historical data using k-nearest-neighbor search and blends them with the direct forecast. The input window supplies recent observations. A separate historical DataFrame, passed through the context_df argument, supplies the retrieval context. Developers can configure how heavily the final prediction weights each pathway. NVIDIA points to a paper, Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains, for details on DARR.

Interpretability

Kumo Forecast includes a model-agnostic interpretability framework that does not modify the underlying forecaster. When enabled, it returns an explanation bundle containing:

• Lag and feature attributions for each forecast horizon
• Semantic-flow diagnostics
• Trajectory-stability diagnostics
• Optional embedding Integrated Gradients

The method is described in a second paper, NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates.

Limitations

The model accepts only numeric temporal data. In DARR mode, the dimensions of the context data must be similar to those of the input. NVIDIA calls for use-case-specific testing and recommends comparing direct and retrieval-enhanced configurations on the same held-out forecasting windows.

Availability and tooling

The API is DataFrame-first with configurable forecast horizons. The Kumo-TS repository includes an example for fine-tuning on a user-supplied CSV. NVIDIA also lists Kumo Forecast in the Predictive Maintenance module, in Early Access, of its Factory Operations (FOX) Blueprint, alongside Kumo Anomaly, a companion diffusion-based anomaly detection model announced the same day.

What this means

Kumo Forecast is a small model by current standards, and its main differentiator is architectural rather than scale. Retrieval over a user's own history lets a frozen, general-purpose encoder adapt to domain-specific patterns without retraining. The bundled attributions address a common obstacle to deploying forecasts in industrial and operational settings.

The announcement includes no benchmark numbers, so there is no public basis yet for comparing it against other pretrained time-series models. The Apache 2.0 license and fine-tuning example make independent evaluation straightforward. Until then, accuracy claims are unverified, and NVIDIA's own guidance to test per use case applies.