NVIDIA releases Kumo Forecast: a 340M open time-series model that checks the past before predicting

• What: NVIDIA's Kumo Forecast — an open multivariate time-series forecasting model, formerly known as NV-Tesseract Forecasting.
• When: Announced October 21, 2026, in an official NVIDIA post on the Hugging Face blog.
• Size: A roughly 340-million-parameter MOMENT-1-large Transformer encoder (kept frozen) plus a trainable forecasting head.
• License: Apache 2.0 — code at NVIDIA/Kumo-TS on GitHub, weights at nvidia/Kumo-Forecast on Hugging Face.
• Retrieval: DARR mode blends direct forecasts with kNN retrieval over a historical DataFrame (example config: alpha 0.2, k=64, temperature 0.05).
• Inputs: Timestamp plus numeric columns as a pandas DataFrame or CSV/JSON; sequence lengths 256, 512, 1024, or 2048; outputs are pandas DataFrames.
• Training: 3 million data points with an 80/20 split per dataset, starting from the Monash Time Series Forecasting Archive (30 datasets); evaluated on ECL, Traffic, ETTh, and ILI.

NVIDIA released Kumo Forecast on October 21 — an open, Apache-2.0 multivariate time-series forecasting model that does something most forecasters don’t: it consults the past before predicting the future, and then explains its reasoning horizon by horizon. A roughly 340-million-parameter pretrained encoder handles pattern recognition, an optional retrieval step digs through your history for similar moments, and an interpretability framework writes up why the model believes what it believes.

That matters because forecasting is the unglamorous engine under demand planning, predictive maintenance, energy load balancing, and anomaly detection — exactly the places where “trust me, the neural net said so” fails hardest. A forecaster that cites its evidence changes the conversation from blind faith to audited prediction. NVIDIA detailed the release in an official post on the Hugging Face blog, with Unite.AI covering the announcement the same day.

Pretrained backbone, retrievable memory

Kumo Forecast — formerly known as NV-Tesseract Forecasting — builds on the MOMENT-1-large Transformer encoder. The base encoder and its embedder stay frozen while only the forecasting head is trained, separating the pretrained representation model from the task-specific prediction layer. For horizons beyond the model’s native capability, the library supports autoregressive extension that extends predictions recursively.

The distinctive part is DARR: Domain-Aware Representation and Retrieval. In this mode the model runs k-nearest-neighbor retrieval over a historical DataFrame you supply through a separate context_df argument, and blends those retrieved patterns with the direct forecast. The blend is adjustable — the repository’s example sets alpha to 0.2, described in a code comment as 20 percent direct forecasting and 80 percent kNN retrieval, with k of 64 and a temperature of 0.05. In plain terms: the model asks “have I seen something like this before?” and weights the answer. One limitation: the dimensions of the retrieval context must be similar to the context of the input.

The underlying idea is documented in the paper Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains, listed in the repository’s references as appearing in the 2025 IEEE International Conference on Big Data proceedings.

Forecasts that show their work

The second differentiator is a model-agnostic interpretability framework that runs without modifying the underlying forecaster. When enabled, each forecast ships with an explanation bundle: lag and feature attributions per forecast horizon, semantic-flow diagnostics, trajectory-stability diagnostics, and optional embedding Integrated Gradients. Pass interpretability=True in the SDK and the bundle is written 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 (arXiv, August 25, 2026), whose authors report competitive or superior faithfulness versus standard post-hoc baselines while being substantially more computationally efficient — along with a stability analysis that flags regimes where interpretation should be applied with caution.

The practical shape: DataFrames in, DataFrames out

The library is deliberately boring in the right ways. The API is DataFrame-first with configurable forecast horizons. Inputs must be numeric temporal data — a timestamp column plus one or more numeric value columns, supplied as a pandas DataFrame or CSV/JSON — with supported input sequence lengths of 256, 512, 1024, or 2048 observations. Outputs are pandas DataFrames containing timestamps and predicted values for the specified horizon.

It 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. The repository declares Python 3.10 or later with PyTorch 2.7 or later, recommends a GPU with automatic fallback to CPU, and auto-downloads pretrained weights from Hugging Face.

Training, per the model card, used 3 million data points with an 80/20 split per dataset, starting with the Monash Time Series Forecasting Archive — described as 30 primary datasets including missing-value variations. Listed evaluation datasets include ECL (electricity consumption from 370 clients, 15-minute intervals, 2011–2014), Traffic (road-occupancy rates from San Francisco Bay Area freeway sensors), ETTh (hourly transformer oil temperature and six power-load features over two years), and ILI (weekly U.S. CDC influenza-like-illness ratios). The model card also states that integration into AI systems requires use-case-specific testing following the V-model methodology — the right caveat for a model aimed at production floors.

Where it lands in NVIDIA’s open stack

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 release sits inside the Kumo Structured open-model family — Kumo Tabular, Kumo Relational, Kumo Time Series, and Kumo Forecast — which NVIDIA folded into its expanded Agent Toolkit at GTC Berlin on October 21, positioning the family as the data-facing layer for enterprise agents.

The first deployment story is industrial: NVIDIA’s Factory Operations Blueprint includes Early Access to a Predictive Maintenance module that pairs Kumo Anomaly with Kumo Forecast to analyze real-time data and predict future anomalies.