NVIDIA Announces Open Kumo Forecast Multivariate Time-Series Model

NVIDIA announced Kumo Forecast on October 21, 2026, an open multivariate time-series forecasting model that combines a pretrained MOMENT-1-large Transformer backbone with optional retrieval over historical data and a model-agnostic interpretability framework. The model, formerly known as NV-Tesseract Forecasting, is available under the Apache 2.0 license, and NVIDIA detailed it in an official post on the Hugging Face blog.

The post was authored by NVIDIA’s Saira Qureshi, Aditi Gautam, Manikandan Ravikiran, Sena Ekiz, and Viraj P Modak. Model code is distributed through the NVIDIA/Kumo-TS GitHub repository, an open-source time-series library that also houses the diffusion-based Kumo-Anomaly package, while pretrained weights ship through the nvidia/Kumo-Forecast model card on Hugging Face. The card lists a Hugging Face release date of 04/07/2026, global deployment, and a stated use case of companies, organizations, and research hubs running long- or short-horizon forecasting on temporal data.

Pretrained Backbone and Direct Forecasting

Kumo Forecast uses the MOMENT-1-large Transformer encoder as its backbone, with an approximately 340-million-parameter base encoder plus a trainable forecasting head. Encoder and embedder weights remain frozen during training while only the forecasting head is trained, separating the pretrained representation model from the task-specific prediction layer. In the direct forecasting path, the head maps the model’s representations to future values for the requested horizon; for horizons beyond the model’s native capability, the library supports autoregressive extension that extends predictions recursively.

The library provides a DataFrame-first API with configurable forecast horizons and optional historical context for retrieval. Inputs must be numeric temporal data consisting of a timestamp column and one or more numeric value columns, supplied as a pandas DataFrame or CSV/JSON, and supported input sequence lengths are 256, 512, 1024, and 2048 observations. Outputs are pandas DataFrames containing timestamps and predicted values for the specified forecast horizon.

DARR Retrieval and Interpretability

Beyond direct forecasting, Kumo Forecast includes Domain-Aware Representation and Retrieval (DARR), a context-enhanced mode that augments predictions with k-nearest-neighbor retrieval from historical data. The SDK accepts that history through a separate context_df argument, creating a distinction between the model’s input window, which supplies recent observations, and the retrieval context, which supplies additional patterns. Developers can adjust how strongly the final prediction relies on each pathway through the library’s configuration options. NVIDIA lists as a limitation that the dimensions of the retrieval context must be similar to the context of the input.

The repository’s DARR example sets an alpha of 0.2, described in a code comment as 20 percent direct forecasting and 80 percent kNN retrieval, together with a k of 64 and a temperature of 0.05. For further detail on the mode, NVIDIA points developers to the paper Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains, which the repository’s reference list places in the 2025 IEEE International Conference on Big Data proceedings.

The interpretability framework is model-agnostic and runs without modifying the underlying forecaster. When enabled, it produces an explanation bundle containing lag and feature attributions per forecast horizon, semantic-flow diagnostics, trajectory-stability diagnostics, and optional embedding Integrated Gradients. Passing interpretability=True in the SDK writes the bundle alongside the forecast as JSON and CSV files plus a PDF report.

The framework follows the preprint NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates, by Muyan Anna Li, Manikandan Ravikiran, and Aditi Gautam, submitted to arXiv on August 25, 2026. The paper models forecasting as a latent trajectory, introduces semantic flow to quantify how information evolves across time in a model’s internal representations, aggregates that flow into a lag-horizon attribution matrix, and applies structure-preserving perturbations with sparse local surrogate models to produce human-readable explanations. The authors report that the semantic-flow variant achieves competitive or superior faithfulness compared with standard post-hoc baselines while being substantially more computationally efficient, and that their stability analysis identifies regimes where interpretation should be applied with caution.

Training Data, Runtime Requirements, and Integrations

According to the model card, training used 3 million data points with an 80/20 split per dataset, starting with the Monash Time Series Forecasting Archive, described as a diverse collection of 30 primary datasets that includes missing-value variations. The card’s listed evaluation datasets comprise ECL, electricity-consumption data from 370 clients recorded every 15 minutes from 2011 to 2014; Traffic, road-occupancy rates measured by sensors on San Francisco Bay area freeways; ETTh, electricity-transformer oil temperature and six power-load features collected hourly over two years; and ILI, weekly U.S. Centers for Disease Control data describing the ratio of patients with influenza-like illness to total patients. The card also cites the MOMENT, TimeRAF, and Beyond MAE temporal-dependence error papers as references.

Kumo-Forecast runs on PyTorch with support for NVIDIA Ampere and Hopper microarchitectures on Linux, and the model card reports inference testing on A100 and H100 systems with eight GPUs each. The repository declares Python 3.10 or later with PyTorch 2.7 or later, pandas, and numpy for Kumo-Forecast, recommends a GPU with automatic fallback to CPU, and auto-downloads pretrained weights from Hugging Face. The model card states that the models are designed and optimized for NVIDIA GPU-accelerated systems, and that integration into AI systems requires additional use-case-specific testing following the V-model methodology.

For application-specific evaluation, NVIDIA advises comparing the direct and retrieval-enhanced configurations on the same held-out forecasting windows and validating forecast quality across the intended horizons and operating conditions. The Kumo-TS repository additionally provides an example for fine-tuning Kumo-Forecast on a user-supplied CSV, giving developers both a pretrained inference workflow and a documented starting point for adapting the model to their own data. On the deployment side, the Factory Operations Blueprint includes Early Access to a Predictive Maintenance module that uses Kumo Anomaly and Kumo Forecast to analyze real-time data and predict future anomalies, and NVIDIA’s post links a GTC San Jose 2026 session covering platform integration.