A 30M Parameter Model Beats Google's AI Giant at Pure Math

According to SiliconANGLE, artificial intelligence startup Synthefy has secured $6.5 million in seed funding to launch a new category of AI architecture known as Structured Data Foundation Models. Rather than parsing sentences, these specialized systems are trained directly on raw numerical tables and time-series data to execute high-stakes enterprise calculations. By shifting focus from language to mathematical structures, the company is targeting a fundamental bottleneck in how modern businesses process financial and operational data.

Funding Round Accelerates Dedicated Numerical Architectures

The $6.5 million seed financing round was led by Wing Venture Capital, with participation from Haystack, Samsung Next, Canonical Crypto, and Lightscape, alongside strategic angel investors from OpenAI, Microsoft, and Meta. Synthefy Inc. aims to redefine enterprise computing by building foundation models designed from the ground up for numerical math rather than text processing.

While traditional large language models process numbers by breaking them into arbitrary text tokens, Synthefy's approach preserves the structural relationships inherent in financial spreadsheets, demand matrices, and sensor logs.

Nori Outperforms Google's TabFM at Two Percent of Its Size

The core breakthrough behind Synthefy's strategy lies in its first open-source model, Nori. Despite containing only 30 million parameters, Nori demonstrates surprising efficiency when processing complex tabular data:

Outperforms Google's 1.6-billion parameter TabFM model while operating at just 2% of its size when thinking features are enabled
Eliminates weeks of manual data preparation and fine-tuning required by traditional machine learning algorithms like XGBoost and LightGBM
Achieved over 600,000 downloads within weeks of its initial unannounced release

Synthefy Chief Executive Officer Somi Agarwal noted in an interview that legacy systems force engineering teams to start from scratch for every new project. «That work does not compound,» Agarwal explained, emphasizing that pre-training models on millions of synthetic datasets allows companies to get instant predictions in minutes rather than months.

What Dedicated Number Crunching Means for Everyday Systems

The broader implications of dedicated numerical models extend far beyond server rooms and data science departments. By replacing bloated multi-billion parameter language models with lean, specialized mathematical engines, enterprises can run hyper-accurate predictions directly on local edge devices and corporate networks with minimal energy usage.

For consumers, this algorithmic shift translates into real-time dynamic pricing that reflects immediate supply changes, vastly improved fraud protection at checkout, and more reliable logistics that prevent supply chain shortages before they disrupt store shelves.

Why it matters

The arrival of Structured Data Foundation Models marks a structural turning point for enterprise technology deployment. For years, organizations attempted to force text-centric large language models into handling complex tabular data, incurring massive compute costs and frequent calculation errors. Synthefy's lightweight approach proves that domain-specific models can deliver superior accuracy at a tiny fraction of the computational footprint. This shift lowers the financial barrier for mid-sized enterprises seeking automated fraud detection, supply chain forecasting, and dynamic pricing. As hardware demands scale globally, adopting targeted numerical architectures will help prevent infrastructure bottlenecks while enabling real-time edge processing for critical financial and industrial operations.

FAQ

What is Synthefy's Nori model?

Nori is an open-source Structured Data Foundation Model with 30 million parameters, fine-tuned specifically to process numerical and tabular data rather than natural language text.

How does Synthefy compare to traditional machine learning models?

Unlike traditional frameworks like XGBoost that require weeks of data preparation and retraining for every new problem, Synthefy's models draw on pre-trained synthetic datasets to generate accurate predictions in minutes.

Who backed Synthefy's seed funding round?

The $6.5 million seed round was led by Wing Venture Capital, with participation from Haystack, Samsung Next, Canonical Crypto, Lightscape, and angel investors from OpenAI, Microsoft, and Meta.