Gretel is a synthetic data generation platform for enterprises and developers, providing differential privacy synthesis capabilities for structured tabular data, time series data and text data. Its open source gretel-synthetics library (PyPI 0.22.20) supports ACTGAN, Timeseries DGAN and other models. It has been acquired by NVIDIA and integrated into the Nemotron/NeMo ecosystem. The original SaaS platform has stopped operating independently.

Gretel’s in-depth review: the panorama and endgame of the synthetic data generation platform

Core parameters and statistics

Dimensions	Specifications
Product Positioning	Differential Privacy Synthetic Data Generation Platform (Structured/Sequential/Text)
Current status	Acquired and integrated by NVIDIA, independent SaaS has ceased operations
Open source license	Gretel Source Available License (main library) / Apache-2.0 (client, etc.)
Supported platforms	Web (service discontinued), API (service discontinued), Python SDK (open source available)
Programming Language	Python (99.7%), supplemented by Shell, Jupyter Notebook
GitHub Stars	685 (gretel-synthetics main repository)
GitHub Forks	101
Number of organization warehouses	30 (all archived)
Last version	v0.22.20 (2025-06-25)
Date of filing	2026-02-18
PyPI downloads	Publicly available
Core Models	ACTGAN, Timeseries DGAN, LSTM (deprecated)

Gretel's independent operating life cycle spanned approximately 7 years (2018–2025), culminating in its acquisition by NVIDIA and integration into the Nemotron/NeMo ecosystem. Its open source assets (especially the gretel-synthetics library) are still available independently, but the SaaS control plane and cloud API have been taken offline with the acquisition. For technical teams evaluating synthetic data tools, Gretel's open source library is still a highly valuable reference for differential privacy synthetic data implementation.

User and market recognition

GitHub Community Base: The gretel-synthetics main library accumulated 685 Stars and 101 Forks before archiving, and 29 contributors participated in code submission, reflecting the attention of developers in the field of synthetic data to the project. There are also auxiliary warehouses such as gretel-blueprints (85 Stars) and gretel-python-client (64 Stars) under the organization.

Industry Consolidation Signal: In early 2026, Gretel was acquired by NVIDIA, and the domain name gretel.ai now redirects to the NVIDIA synthetic data solution page. This signifies that NVIDIA regards Gretel's synthetic data capabilities as a key puzzle piece of its AI Agent data infrastructure, forming a complete data link with the Nemotron model NeMo fine-tuning platform. The acquisition itself is the highest endorsement of the value of Gretel’s technology.

Enterprise customers and financing: Gretel served the synthetic data needs of many financial, medical and technology companies during its independent operation, but the specific customer list and financing round amount have not been fully disclosed. After the acquisition, its enterprise customers have migrated to the NVIDIA ecosystem.

Cost advantage

C side/individual users: Gretel’s open source Python library gretel-synthetics remains publicly available on PyPI, and individual developers can install and use it at zero cost. The free quota of the original SaaS platform (such as the number of free records generated every month) has been terminated when the platform went offline.

Developers/API Users: The original Gretel REST API (called via gretel-python-client) is no longer available stand-alone. For developers who want to consume synthetic data capabilities through APIs, they need to turn to the synthetic data service (NeMo Data Designer) of the NVIDIA Nemotron platform. The pricing model is based on the NVIDIA official pricing page.

Enterprise/Private Deployment: The open source library supports localized deployment to run synthetic data generation on your own infrastructure without the need for a network connection. Enterprises can build their own pipelines based on gretel-synthetics, and the cost is limited to hardware resources (GPU recommended) and operation and maintenance manpower. NVIDIA Enterprise's synthetic data solutions require licensing through NVIDIA commercial channels.

Main functions

• Structured tabular data synthesis: Based on the ACTGAN (Anyway CTGAN) model, it can learn column distribution and correlation from real tabular data and generate synthetic data sets with high statistical fidelity. Supports enterprise-level features such as automatic column type detection, memory-optimized NaN proxy handling, and more. Suitable for database desensitization, test data generation, data enhancement and other scenarios.

• Time series data synthesis: Based on Timeseries DGAN (PyTorch implementation of DoppelGANger), it is specially designed for time-dependent data such as smart home sensor readings and financial time series IoT device logs. Preserve autocorrelation and cross-feature dependencies of time series data.

• Differential Privacy Protection: Core differentiation capabilities - not only can "look like real data", but also provide quantifiable privacy protection guarantees through the Differential Privacy (Differential Privacy) mechanism. Users can control the privacy budget through parameters and make a trade-off between data utility and privacy protection intensity.

• Data Quality and Privacy Report: After generating synthetic data, Gretel automatically generates data quality reports (covering column distribution comparisons, correlation matrices, field missing rates, etc.) and privacy assessment reports (privacy risk scores, member inference risk assessments) to help users verify whether the synthetic data achieves the expected effectiveness.

• Code Synth (Code Data Synthesis): Experimental direction—expanding training data for AI programming models through code synthesis. The gretel-code-synth repository has explored this direction, but only in the prototype stage.

Expert View: The core value of Gretel does not lie in the single function of "generating data", but in the combination of "privacy protection + data utility assessment" - it not only generates synthetic data, but also quantifies the performance of synthetic data in the risk of privacy leakage. This "generation-evaluation" collaborative mechanism has practical engineering significance in strong regulatory scenarios such as financial compliance and medical data sharing. After being integrated by NVIDIA, this capability has been integrated into the Safe Synthesizer module of the Nemotron platform, and collaborates with the visual design tool of NeMo Data Designer to lower the threshold for using synthetic data.

Model and version evolution

Open source library development phase (2019–2025)

Gretel's core open source library gretel-synthetics has gone through a complete iteration cycle, evolving from the initial TensorFlow-based LSTM model to a comprehensive library that supports multiple generation architectures.

Early version (v0.1–v0.17): Mainly based on the LSTM model based on TensorFlow, supporting structured text synthesis and laying the foundation for the infrastructure.

Interim version (v0.18–v0.21): Introducing ACTGAN (CTGAN extension based on SDV) to support tabular data; adding Timeseries DGAN (based on DoppelGANger) to cover time series scenarios. This phase witnessed the transformation from a single LSTM to a multi-model strategy.

Later version (v0.22.x, 2024–2025): Maintenance period, focusing on dependency security updates (jinja2, axios, TensorFlow compatibility), Python version upgrade (to 3.11.9), and removal of the deprecated LSTM runtime. v0.22.20 is the final version.

Version Timeline

Version	Date	Critical Changes
v0.22.20	2025-06-25	Python 3.11.9, remove TensorFlow dependency, depend on security update
v0.22.19	2025-02-26	Remove LSTM runtime, lower timing test assertion threshold
v0.22.17	2025-01-29	Fix document build dependency
v0.22.16	2024-12-11	Dependency and compatibility maintenance
v0.22.15	2024-12-05	Fix TensorFlow/protobuf dependency
v0.22.14	2024-10-30	Remove Slack notification integration
v0.22.13	2024-10-09	TensorFlow 2.12.1 Update
v0.22.12	2024-09-11	General fixes
v0.22.11	2024-06-27	General fixes
v0.22.10	2024-05-15	NaN proxy for correlation calculations

Post-acquisition phase (2026–present)

On February 18, 2026, Gretel's GitHub organization was marked as archived by the administrator, and all repositories became read-only. The gretel.ai domain started redirecting to NVIDIA. Gretel's technical capabilities, specifically differentially private synthetic data generation, are integrated into the NeMo Data Designer and Safe Synthesizer components of the NVIDIA Nemotron platform. Original SaaS users have been migrated to the NVIDIA ecosystem.

Technical advantages

Theoretical basis of privacy protection: Gretel's core technical difference lies in the engineering integration of differential privacy (Differential Privacy) theory into the synthetic data generation process. By injecting controlled noise into the training process, the difficulty of inferring original individual records from synthetic data is theoretically quantifiable. This means that data scientists can accurately answer "What is the probability that this synthetic data set leaks real information about a user?"

Mechanics → Effects → Scenarios:

• Mechanism: Apply differential privacy perturbation (through algorithms such as DP-SGD) to the gradient or loss function during the model training stage instead of post-processing desensitization after generation.
• Effect: Synthetic data retains the original data distribution characteristics (mean, variance, correlation) at the statistical level, but provides privacy protection guarantee at the individual record level.
• Scenario: When medical research institutions share patient data, they can use Gretel to generate privacy-protected data sets for use by external partners; financial institutions can use synthetic data to replace sensitive transaction records in model development.

Engineering wisdom for multi-model strategies: Gretel does not rely on a single generation algorithm, but provides three models: ACTGAN (tabular data), Timeseries DGAN (time series data) and LSTM (text, deprecated). This design of "selecting models by data type" avoids the compromise in performance caused by the "universal model". In actual use, it is necessary to select an appropriate model based on the data form.

The cost of differential privacy: The addition of differential privacy will inevitably reduce the statistical fidelity of synthetic data. The smaller the privacy budget $\epsilon$ (the stronger the protection), the greater the data utility loss. Gretel allows users to choose a balance point on the privacy-utility curve, but engineering teams need to evaluate the tolerance of data accuracy for specific business scenarios.

How to use

Method 1: Use open source Python library (currently available)

#Install gretel-synthetics
pip install gretel-synthetics

# Note: dependencies need to be installed manually
pip install torch==2.0 # for Timeseries DGAN or ACTGAN (via SDV)
# or
pip install sdv<0.18 # for ACTGAN

Basic usage process:

1. Prepare training data (Pandas DataFrame)
2. Configure training parameters (select model, privacy budget, etc.)
3. Training the synthetic model
4. Generate synthetic records
5. Evaluate synthetic data quality

Detailed tutorials can be found in the examples/ directory of the GitHub repository with Jupyter Notebook examples, or run online via Google Colab.

Method 2: NVIDIA Nemotron Ecosystem (currently recommended)

Gretel's capabilities are integrated into NVIDIA NeMo Data Designer:

1. Visit https://build.nvidia.com/nemo/data-designer
2. Configure seed data set and synthetic columns
3. Select privacy protection level
4. Preview and batch generate

This method is suitable for enterprise users who require enterprise-level SLA, visual operation interface and large-scale production.

Method three: NVIDIA Safe Synthesizer (privacy compliance scenario)

For data generation scenarios that require HIPAA/GDPR compliance, use the Safe Synthesizer component in the NVIDIA Nemotron platform, which inherits and extends Gretel's differential privacy capabilities.

Product Pricing

Tiers	Pricing Model	Description
Open source library (self-hosted)	Free	gretel-synthetics PyPI package, completely free to use, no usage limit
NVIDIA NeMo Data Designer	Subject to NVIDIA official pricing	Provides visual designer, large-scale generation, enterprise-level support
NVIDIA Enterprise License	Contact NVIDIA Business	Includes SLA, compliance certification, private deployment options

Gretel's original SaaS pricing (based on record-count subscriptions) has been taken offline with the platform and is no longer available. The current synthetic data capabilities are delivered through the NVIDIA ecosystem. The specific price is subject to real-time information on NVIDIA's official pricing page. It is recommended that enterprise users confirm the business terms with the NVIDIA sales team.

Application scenarios

• Data desensitization and compliance sharing: Financial institutions need to share transaction data with external risk control model developers, but are subject to privacy regulations such as GDPR/CCPA. Using Gretel's ACTGAN to generate synthetic transaction data with similar statistical characteristics, combined with differential privacy protection, can not only meet the needs of model development, but also pass compliance audits. Verification focus: Whether the privacy budget $\epsilon$ is set to meet legal requirements, and whether the synthetic data retains the original characteristics in key indicators (such as fraud rate distribution).

• AI Agent training data generation: When building conversational AI agents, real conversation data is often scarce and sensitive. Gretel’s (now NVIDIA Nemotron) synthetic data capabilities can generate large-scale diverse conversation datasets from a small number of seed conversations, covering edge scenarios and intent variations. Key points of verification: Whether the diversity and naturalness of synthetic dialogue are enough to cover the long-tail distribution in real dialogue.

• Machine Learning Testing and CI/CD: The continuous integration pipeline requires a large amount of test data to verify the correctness of the data processing logic. Gretel can generate structurally consistent synthetic data sets from the schema of the production database for unit testing and integration testing, preventing production data from being copied to the development environment. Key points of verification: Whether the synthetic data completely covers data constraints (unique keys, foreign keys, enumeration value ranges, etc.).

• Medical Data Research: When hospitals collaborate with research institutions, patient data is strictly protected by HIPAA. Gretel's differentially private synthetic data can generate synthetic patient records with consistent statistical characteristics without exposing individual patient information, supporting scenarios such as epidemiological research and treatment effect analysis. Verification focus: Whether the addition of differential privacy leads to the loss of rare disease characteristics and affects the validity of research conclusions.

Applicable people

• Data Privacy and Compliance Engineer: Need to provide "usable but non-traceable" data sets to third parties without violating privacy regulations. Gretel’s Differential Privacy Report and Privacy Score provide auditable evidence of compliance. However, please note: Differential privacy cannot completely eliminate the risk of re-identification, and the compliance team still needs to work with legal counsel to evaluate specific scenarios.

• Data Scientists and ML Engineers: Use synthetic data for data enhancement when training data is insufficient or data distribution is unbalanced. Gretel’s multi-model strategy allows data scientists to choose the most appropriate generation algorithm based on data type (tabular/time series/text). Unsuitable scenarios: In scenarios where the authenticity of synthetic data is extremely high (such as directly using synthetic data to train a production model), it is recommended to first use small-scale experiments to verify the improvement of the model effect of synthetic data.

• DevOps/QA Engineer: A large amount of test data is required to verify the correctness and boundary conditions of the data processing pipeline. Gretel's open source library can be embedded into CI/CD pipelines to generate test data on demand. But please note: open source libraries do not provide data constraint integrity guarantees (such as foreign key consistency), and the synthesis of complex schemas requires additional engineering work.

• Enterprise Technology Selection Decision Makers: When evaluating synthetic data tools, Gretel's open source legacy (gretel-synthetics) is the reference implementation of choice for technical validation. However, production-oriented large-scale synthetic data solutions should directly evaluate the complete capabilities of the NVIDIA Nemotron platform. Not suitable for the boundary: If you need a purely managed SaaS service that does not rely on the NVIDIA ecosystem, Gretel is no longer suitable - the original SaaS has been offline and you need to turn to other independent synthetic data vendors.

Summary and Outlook

As a pioneer in the field of synthetic data generation, Gretel's core contribution lies in combining differential privacy theory with engineering practice to create a complete data synthesis workflow of "generation-evaluation-compliance". Its open source library gretel-synthetics provides the industry with a verifiable baseline implementation of privacy-preserving synthetic data.

Acquired by NVIDIA and integrated into the Nemotron/NeMo ecosystem, it marks the industry trend of synthetic data moving from independent tools to platform infrastructure. Gretel's technical DNA has now become part of NVIDIA's AI data strategy, forming in-depth synergy with the Nemotron model NeMo fine-tuning platform LangChain Agent framework.

Current Limitations and Uncertainties:

• The independent SaaS has been offline, and the original user experience (Web UI, API Key management, team collaboration) cannot be reproduced.
• Although the open source library is available, the long-term maintenance status is unknown (the warehouse has been archived, and there is no obvious signal of active community fork takeover).
• Within the NVIDIA ecosystem, there is no public roadmap as to whether Gretel's differential differentiation capabilities are fully integrated with NVIDIA's original product lines (such as NeMo Safe Synthesizer) and whether an independent brand will remain.

Procurement/Adoption Risk Assessment:

• Technical verification: The open source library of gretel-synthetics is suitable for technical concept proof of differential privacy synthetic data, and can be started at zero cost.
• Production use: It is recommended to directly evaluate the synthetic data service (NeMo Data Designer + Safe Synthesizer) of the NVIDIA Nemotron platform to confirm whether it meets privacy compliance and data scale SLA requirements.
• Migration costs: If you already have existing assets on Gretel SaaS, you need to contact NVIDIA to confirm the migration plan; if you start a new project, it is recommended to use the NVIDIA ecosystem as the starting point for evaluation by default.
• Vendor lock-in: Choosing the NVIDIA ecosystem means accepting its platform dependency, and it is recommended to also evaluate open source alternatives (such as SDV, YData Synthetic) as alternatives during the PoC stage.

Version Info

• gretel-synthetics v0.22.20 ：Updated axios and jinja2 dependencies, removed TensorFlow dependencies, upgraded Python to 3.11.9, and lowered the timing DGAN test assertion threshold. This is the last public version of Gretel during its independent operation, and was archived in the GitHub repository on 2026-02-18. （2025-06-25）
• gretel-synthetics v0.22.19 ：Downgrade setuptools_scm, remove LSTM runtime, fix MM Report. There is no official precise date yet. （2025-02-26）
• gretel-synthetics v0.22.17 ：Fix documentation build, update jinja2 to 3.1.5. There is no official precise date yet. （2025-01-29）
• gretel-synthetics v0.22.10 ：Replaced NaN proxy for correlation calculations, updated jinja2 dependency. There is no official precise date yet. （2024-05-15）