Kumo-Anomaly Overview

Description:

NVIDIA Kumo-Anomaly provides anomaly detection functionality. Rather than relying only on transformers, it introduces diffusion modeling, stabilized through curriculum learning, and pairs it with adaptive thresholding methods in a model purpose-built for anomaly detection. Together, these elements address some of the most challenging issues in the field: noisy, high-dimensional signals that drift over time and contain rare, irregular events.

License/Terms of Use:

Governing Terms: Model: The model is provided under the Apache License, Version 2.0.

Deployment Geography:

Global

Use Case:

Companies, organizations, research hubs looking to do anomaly detection on temporal data.

Release Date:

29th June 2026 via https://huggingface.co/nvidia/Kumo-Anomaly

Reference(s):

Segmented Confidence Sequences and Multi-Scale Adaptive Confidence Segments for Anomaly Detection in Nonstationary Time Series
NV-Tesseract-AD: Curriculum-Enhanced Conditional Diffusion for Multivariate Time-Series Anomaly Detection

Model Architecture:

Architecture Type: Diffusion based transformer

Network Architecture: ResNet34

Number of model parameters: 8 million

The Kumo-Anomaly is a diffusion-based model for time-series imputation and anomaly detection.

Input:

Input Type(s): Tabular numeric
Input Format(s): Tabular Pandas DataFrame or CSV/JSON
Input Parameters: Two-Dimensional (2D)
Other Properties Related to Input: Pre-Processing Needed
Anomaly detection: Contains timestamp column and one or more numeric value columns.

Output:

Output Type(s): Tabular numeric
Output Format: Tabular Pandas DataFrame
Output Parameters: Two-Dimensional (2D)
Other Properties Related to Output: Post-Processing Needed
Anomaly detection: Contains timestamp, value, and anomaly label columns.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration:

Runtime Engine(s):
PyTorch

Supported Hardware Microarchitecture Compatibility:
NVIDIA Ampere
NVIDIA Hopper

[Preferred/Supported] Operating System(s):
Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s):

Kumo-Anomaly

Training & Testing Datasets:

80/20 split per dataset.

Data Modality: Other: Numeric time series

Training Data Size: 3 million data points

TSB-AD-M

Data Collection Method by dataset
Hybrid: Automatic/Sensors, Human, Synthetic
Labeling Method by dataset
Hybrid: Human, Automatic, Synthetic
Properties: The TSB-AD-M benchmark consists of 17 public multivariate datasets containing 198 curated multivariate time series spanning diverse application domains. The datasets include both point and sequence anomalies with varying dimensionalities, anomaly ratios, anomaly lengths, and time-series lengths, providing a comprehensive benchmark for multivariate time-series anomaly detection.

Evaluation Datasets:

Detecting Anomalies in Wafer Manufacturing

Data Collection Method by dataset
Automatic/Sensors
Labeling Method by dataset
Human, Automatic
Properties: High-dimensional tabular anomaly detection dataset collected from semiconductor wafer manufacturing. Each sample represents a single wafer described by 1,558 anonymized process features and a binary anomaly label indicating normal or anomalous production.

CalIt2 Building People Counts

Data Collection Method by dataset
Automatic/Sensors
Labeling Method by dataset
Human
Properties: Time-series dataset containing people counts collected from sensors at the main entrance of the Calit2 building at the University of California, Irvine. Measurements were recorded every 30 minutes over approximately 15 weeks, producing 48 observations per day.

Network Traffic

Data Collection Method by dataset
Synthetic, Automatic/Sensors
Labeling Method by dataset
Automatic, Human
Properties: Labeled network traffic dataset for anomaly detection containing normal and malicious traffic records represented by multiple network features. The dataset is intended for evaluating supervised anomaly detection methods on network intrusion scenarios.

Genesis Demonstrator

Data Collection Method by dataset
Automatic/Sensors
Labeling Method by dataset
Human
Properties: Multivariate industrial time-series dataset collected from the Genesis portable pick-and-place demonstrator developed as part of the OPAK and IMPROVE projects. The dataset contains continuous and discrete sensor signals together with timestamps and labeled anomalous operating conditions.

Falling People

Data Collection Method by dataset
Human, Automatic/Sensors
Labeling Method by dataset
Human
Properties: This data set was used during a thesis to develop safer smart environments. The origin is a care independent smart home environment to detect the falling of elderly people.

Inference:

Engine: PyTorch, Transformer Engine
Test Hardware:

• A100 (8 GPUs; each is 74 GB)
  
• H100 (8 GPUs; each is 80 GB)

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.