Synthefy raises $6.5M for numerical data foundation models

Synthefy describes itself as the pioneer of what it calls "Structured Data Foundation Models," or SDFMs. Similar to how LLMs learn to reuse word patterns by digesting vast amounts of text-based data, Synthefy's SDFMs use numerical data to learn about numbers and calculations. By ingesting vast amounts of number-crunching data, the models can preserve the intricate relationships within time-series data and tables, enabling them to generalize across numerical data and perform calculations faster and more accurately than standard machine learning models, the company says.

Alongside the funding announcement, Synthefy discussed its first open-source SDFM, Nori, which was quietly released a few weeks ago. Nori has an extremely lightweight architecture, but despite this it packs strong performance. In testing, a 30-million-parameter version of Nori outperformed Google LLC's 1.6-billion-parameter TabFM model. It performs even better when its "Thinking" capabilities are enabled, surpassing Google's model despite being just 2% of its size, the company said.

Nori and other SDFMs are designed to tackle workloads such as fraud detection and dynamic pricing. Traditionally, enterprises use machine learning frameworks such as LightGBM and XGBoost for these tasks. But Synthefy co-founder and CEO Somi Agarwal told SiliconANGLE that most enterprises have to spend weeks on data preparation, training and fine-tuning to create models that can perform reliably enough.

"That work does not compound," he said. "Each new fraud, pricing or forecasting problem starts again. But with Nori, a team can point the model at a new table or problem and get a strong prediction without training or tuning a new model for that dataset. That can bring the initial evaluation down from weeks or months to just minutes."

According to Agarwal, SDFMs can be further enhanced when pretrained on specific tasks. "Nori has learned from millions of synthetic datasets, so it comes to a new problem with experience rather than starting from zero," he said. "Even a small improvement in pricing, fraud or demand prediction can be worth a great deal at enterprise scales."

Synthefy offers access to its capabilities through an open model, a managed application programming interface or deployments within customers' own computing environments. Looking ahead, the company will use the funding to accelerate research, hire more engineers and build the next generation of Nori. It also is open to new industry partnerships. The first version of Nori has been downloaded more than 600,000 times, just weeks after being made available, according to the report.

Despite open-sourcing its models, Synthefy believes it can make money by offering premium support and capabilities atop its basic foundation layer. "The commercial opportunity is the enterprise product around the model, such as managed API usage, proprietary features, private deployments, security and governance controls, integrations, support and large-scale production infrastructure," Agarwal said.

Wing Venture Capital founding partner Gaurav Garg said he is betting on Synthefy's structured data models to become the next major expansion of the AI model market. "Synthefy is building a model platform that can address some of the largest and most valuable datasets in the world," he explained. "Its combination of technical performance, efficient architecture, open models and early enterprise traction positions the company to define this emerging category."