Mostly AI is an open source synthetic data platform originated in Austria, providing two delivery forms: SDK and data intelligence platform. The core is based on the TabularARGN deep generation architecture and supports the training, generation and quality verification of table and text data. It was acquired by Syntho in June 2026 and now operates under the brand "MOSTLY AI, powered by Syntho".

MostlyAI’s synthetic data platform

Core parameters and statistics

Parameter	Value
Product positioning	Open source synthetic data generation platform
Delivery form	Python SDK (open source) + data intelligence platform (enterprise level)
Open Source License	Apache 2.0 (SDK)
Core Architecture	TabularARGN (Deep Generative Model)
Supported data types	Tables (single table/multiple tables/time series), text
Supported platforms	Web (platform), API, desktop (SDK runs locally)
GitHub Stars	~784
GitHub Forks	~68
Total Releases	116 Releases
Latest version	6.1.1 (2026-05-08)
Place of Belonging	Austria (AT)
Current Operator	Syntho (Netherlands, acquired June 2026)

Mostly AI takes "democratization of synthetic data" as its core concept, extending data generation capabilities from enterprise data scientists to ordinary developers. The open source SDK provides a native-first Python tool chain, while the enterprise platform provides visual data exploration, collaboration and sharing capabilities. The two forms cover all scenarios from personal experiments to enterprise-level compliance deployment.

User and market recognition

• Enterprise Customers: Open cases include Swiss Post (increasing customer data availability from 11% to 100%), Erste Group (accelerating model development with synthetic data in the Databricks context), AWS (for data unlocking in customer cloud migrations), Databricks (for Clean Room cross-industry data collaboration).
• Finance & Insurance: JPMorgan uses synthetic data sandbox to accelerate third-party POC; Anthem uses synthetic data to detect fraud and personalize services; Telefónica uses synthetic customer data for analysis.
• Open Source Community: GitHub 784 Stars, 68 Forks, 116 Releases, 20 contributors, PyPI downloads continue to grow. The project operates under the Apache 2.0 license and accepts community contributions.
• Industry Recognition: Listed as a representative vendor of synthetic data by Gartner. Gartner predicts that “by 2030, the majority of data used for the development of AI and analytics projects will be synthetically generated.”

Cost advantage

C-side/individual developers: The open source SDK is completely free (Apache 2.0 license), and can be installed through pip/uv mostlyai[local] to run locally on the local CPU/GPU. There are no explicit costs for personal learning, experimentation, and prototype development, and you only need to bear the cost of local computing resources.

API/Developer: The SDK supports LOCAL mode (local training inference) and CLIENT mode (connecting to remote SDK endpoints). The specific billing methods for CLIENT mode and platform API are not disclosed. You need to contact Syntho Business to obtain a quote. The PyPI package itself is free, but using remote endpoints involves platform resource consumption.

Enterprise/Private: Enterprise-grade Data Intelligence Platform supports Kubernetes/OpenShift privatized deployments, with costs based on data volume, number of nodes, number of users, and support level. There is no official public price list, and the business contract shall prevail. The open source SDK can also be deployed offline on the corporate intranet, but downloading the Hugging Face model requires manual file transfer in an offline environment.

Hidden Cost: Training large language models requires GPU resources (officially recommends modern GPUs for LLM fine-tuning inference); tabular data training can run efficiently on the CPU. Enterprise deployments need to maintain their own Kubernetes infrastructure.

Main functions

• AI generates synthetic data: Based on the TabularARGN deep generation model, it learns statistical distribution and correlation structure from real sample data, and generates synthetic data with consistent statistical characteristics. Supports single table, multi-table relationships, time series data and text data. A quality assurance report (Model Insight Report) is automatically attached to verify fidelity and privacy security.

• AI Assistant: A natural language-driven data analysis assistant that can perform exploratory analysis, trend monitoring, and pattern discovery on production data without writing code. The bottom layer automatically generates and executes Python code, which is suitable for business analysts.

• Mock Data Simulated Data: Generate simulated data based on rules and randomness without the need for real samples. Suitable for the development of contextual demo demonstrations, prototype testing and other scenarios. Mock Data can generate more realistic contextual data than pure rule tools.

• Simulated Data scenario deduction: Model based on real or synthetic data to deduce edge cases, future scenarios and "what-if" conditions. Suitable for stress testing, strategy verification and algorithm robustness evaluation.

• Synthetic Data SDK: Python toolkit, providing four core APIs: mostly.train(), mostly.generate(), mostly.probe(), mostly.connect(). Supports advanced options such as GPU/CPU training, differential privacy, conditional sampling, fairness control, temperature adjustment, etc. Generators can be exported and shared within the platform.

Model and version evolution

Mainline iteration

Mostly AI's open source SDK has undergone an architectural evolution from platform-bound to native-first since it was first released in 2023. Here are the major release milestones:

• 2023 - v1.0 (initial open source): The first open source version based on the TabularARGN architecture, supporting single-table/multi-table synthetic data generation and quality reporting. Licensed under the Apache 2.0 license.

• 2025 Q4 - v5.9.0: Added OAuth2 integration support and improved connectivity to external data sources.

• 2026 Q1 - v5.10.0: Introduce the business rule constraint system (Inequality, FixedCombinations) so that the generated data can comply with specific inter-column relationship constraints. Upgraded engine to 2.4, PyTorch 2.9.1.

• 2026 Q2 - v6.0.0: Major architectural refactor - SDK completely native-first, removing platform-only interfaces and Hive/Databricks/Redshift connectors. CLIENT mode is retained as a standalone lightweight installation option.

• 2026 Q2 - v6.1.0/6.1.1: The engine is iterated to 2.6.1, and the QA module is updated to 1.10.6 to continue to improve training efficiency and generation quality.

Technical White Paper

The arXiv paper "Democratizing Tabular Data Access with an Open-Source Synthetic-Data SDK" (arXiv:2508.00718) will be released in 2025, which systematically explains the SDK architecture TabularARGN model design and quality assessment system.

Technical advantages

TabularARGN Architecture: Mostly AI’s core technical advantage lies in the TabularARGN (arXiv:2501.12012) deep generative model. This architecture is 1-2 orders of magnitude more training efficient than similar models while maintaining the best synthetic data quality. This allows training and generation of millions of records to be completed within minutes even in a CPU-intensive environment—not by compressing the model size, but by optimizing the computational efficiency of the training algorithm.

Flexible multi-model support: Tabular ARGN is used for tabular data, DNN matching is used for graph relationships, and basic LSTM (LSTMFromScratch-3m) or any AutoModelForCausalLM model on fine-tuned Hugging Face Hub can be used for text data. This layering strategy ensures that different types of data use the most appropriate generation method.

Built-in Quality Assurance: Each generated Generator automatically comes with the Model Insight Report, which provides visual reports on fidelity and privacy indicators. Prevent overfitting and leakage when generating, rather than detecting it afterwards.

Advanced Sampling Capabilities: Supports upsampling to any amount of data, conditional simulation based on any column (such as "generate 10,000 24-year-old male records"), under-represented group weight balancing, context-aware data interpolation, statistical fairness control, and temperature regulation rule constraints. These capabilities make synthetic data not just a "copy" but a targeted data enhancement.

Differential Privacy: Differential privacy protection can be enabled during training to provide quantifiable privacy budget guarantee for data release.

How to use

Open source SDK to get started quickly

# Install
# pip install -U 'mostlyai[local]'

from mostlyai.sdk import MostlyAI

#Initialize SDK (LOCAL mode)
mostly = MostlyAI()

# 1. Import pre-trained generator from file
g = mostly.generators.import_from_file(
    "https://github.com/mostly-ai/public-demo-data/raw/dev/census/census-generator.zip"
)

# 2. Detection generates 1000 representative samples
df = mostly.probe(g, size=1000)

# 3. Generate based on fixed column values
df = mostly.probe(g, seed=[{"age": 24, "sex": "Male"}] * 10_000)

# 4. Train your own generator
import pandas as pd
original_df = pd.read_csv(
    "https://github.com/mostly-ai/public-demo-data/raw/dev/titanic/titanic.csv"
)
g = mostly.train(name="Titanic Demo", data=original_df)

# 5. View quality report
g.reports(display=True)

# 6. Generate synthetic data set
sd = mostly.generate(g)
df = sd.data()

Enterprise Platform

Data Intelligence Platform is accessed through a browser and provides:

• Data source connection: direct connection to data platforms such as Databricks, Snowflake, and Postgres
• AI Assistant: Natural language query and analysis
• Project Collaboration: Teams share generators, datasets and analysis reports
• Secure Deployment: Supports Kubernetes/OpenShift privatized deployment

Three operating modes

Mode	Installation method	Applicable scenarios
LOCAL (default)	mostlyai[local] (including PyTorch)	Local development, personal experiments, offline context
CLIENT	mostlyai (lightweight)	Connect to remote SDK endpoint
Docker	Dockerfile build	Isolated context CI/CD integration

Product Pricing

Mostly AI adopts a hybrid pricing model of open source + enterprise platform:

• Open Source SDK: completely free, Apache 2.0 licensed. Can be used locally for unlimited use without functional limitations.
• Enterprise Data Intelligence Platform: Billed on a private deployment or SaaS subscription basis. The specific price has not been disclosed, please contact Syntho Business ([email protected]) for a quote. Pricing factors include data volume, number of users, deployment nodes and support level.
• CLIENT mode: Platform resource consumption is involved when connecting to remote endpoints, and the billing method is not disclosed.

Note: After the acquisition of Syntho in June 2026, the product roadmap and pricing strategy may be adjusted, please refer to the official real-time page.

Application scenarios

• AI/ML Development and Training: Generate high-fidelity training data for machine learning models, solving the problems of scarcity of real data, imbalanced labels, or limited privacy. It can upsample underrepresented populations and correct data bias. Implementation Tips: First use a small sample to do Crawl/Walk/Run progressive training, and then expand after verifying the generation quality.

• Test Data and QA: Generate realistic structured test data for development/test/pre-release context, eliminating the risk of sensitive information leakage of production data. Suitable for integration testing, stress testing, and end-to-end acceptance. Implementation Tips: Mock Data is suitable for scenarios without strict statistical requirements, and synthetic data is suitable for scenarios that require real data distribution.

• Data Sharing and Collaboration: Generate privacy-secure synthetic datasets and share data across teams and organizations without exposing PII. Supports secure data collaboration in Clean Room context. Implementation Tip: Enabling the differential privacy function can provide quantifiable privacy protection and is suitable for compliance audit scenarios.

• Demo and Prototype Display: Generate realistic but non-sensitive data sets for sales demonstrations, product prototypes, and proof-of-concepts to accelerate customer communication and feedback cycles.

• Data enhancement and bias correction: Supplement edge cases or under-represented groups on the basis of original data to improve the generalization ability and fairness of the model.

Applicable people

• Data Scientists and ML Engineers: Use the SDK to generate high-quality synthetic data directly in your Python workflow, without leaving the Notebook environment. Suitable for teams that need to iterate training data frequently.

• Data Engineers and DevOps: Use Docker or CLI mode to embed synthetic data generation into CI/CD pipelines to automatically provision security data for test contexts. Not suitable for real-time scenarios that have extreme requirements on data generation delay.

• Business Analysts & Compliance Teams: Explore data in natural language with the platform’s AI Assistant to gain analytical insights without writing code. Not suitable for scenarios that require deep customization of generated models.

• Enterprise Data Platform Teams: Organizations that need to share data across teams but are constrained by privacy compliance can use Mostly AI to create a synthetic data sandbox. Prerequisite: Kubernetes operation and maintenance capabilities are required.

• Not applicable to people:

  • Scenarios that require unstructured synthetic data such as images, audio, and video (Mostly AI focuses on tables and text)
  • Scenarios with extreme real-time requirements (sub-second level) for generated data
  • Scenarios that do not require statistical authenticity and only require completely random placeholder data (the rule mock tool is more lightweight)

Summary and Outlook

Mostly AI's core competitiveness lies in the efficiency and open source strategy of the TabularARGN architecture - it is one of the few synthetic data platforms that provides Apache 2.0 licenses and has training efficiency 1-2 orders of magnitude higher than competing products. Corporate cases from Swiss Post to JPMorgan have verified its ability to implement in highly regulated industries such as finance, telecommunications, and medical care.

Current Limitations and Uncertainties:

• After being acquired by Syntho in June 2026, the product integration direction, brand strategy and pricing model may change. Enterprises need to pay attention to Syntho's subsequent product announcements before purchasing.
• The pricing of the enterprise platform is not public. You need to contact the business to obtain a contract quotation. It is difficult to directly compare the cost of competing products.
• Text data generation relies on the Hugging Face model, and deployment in offline/air-gapped environments requires additional manual model transmission steps.
• Unstructured data such as images/audio/videos are not supported.

Purchase/Adoption Risk Assessment: Before purchasing the Mostly AI platform, enterprises need to focus on verification: ① The open source strategy of the existing SDK after Syntho's acquisition (whether there are closed source risks); ② The actual deployment cost and minimum resource requirements of the enterprise version; ③ The quality of multilingual data generation such as Chinese; ④ Compliance certification for privatized deployment (such as SOC2, GDPR compliance). It is recommended to first use the open source SDK to do POC to verify the quality of the production, and then conduct business negotiations for the enterprise version after confirming that it meets business needs.

Version Info

• SDK Release 6.1.1 ：Updated engine 2.6.1 and QA 1.10.6, upgraded Dependabot lock files. （2026-05-08）
• SDK Release 6.1.0 ：Updated engine 2.6.0 and QA 1.10.5. （2026-04-30）
• SDK Release 6.0.0 ：SDK native-first refactoring, removing platform-only interfaces and Hive/Databricks/Redshift connectors. （2026-04-27）
• SDK Release 5.10.0 ：Introduce business rule constraints (Inequality and FixedCombinations), upgrade engine 2.4 and PyTorch 2.9.1. （2026-01-08）
• SDK Release 5.9.0 ：Added OAuth2 integration support. （2025-12-12）
• Initial Open Source Release ：The first open source version of the SDK is based on the TabularARGN architecture and supports single-table/multi-table synthetic data generation. There is no official precise date yet. （~2023）