Exclusive: Synthefy raises $6.5M for its number-crunching models trained on numerical data instead of words

There's a version of the AI boom that everyone already knows: massive language models trained on oceans of text, learning to predict the next word. Synthefy Inc. thinks there's a parallel opportunity nobody's built properly yet — the same trick, but for numbers. The startup announced a $6.5 million seed round today, led by Wing Venture Capital with Haystack, Samsung Next, Canonical Crypto and Lightscape joining in, plus angel money from people at OpenAI, Microsoft and Meta.

The pitch is what Synthefy calls Structured Data Foundation Models, or SDFMs. Instead of ingesting text, these models train on tables and time-series data, learning the relationships buried in that structure so they can generalize to new numerical problems rather than starting from scratch each time. The company's first open release, a lightweight model called Nori, quietly went live a few weeks back. A 30-million parameter version reportedly outperformed Google's TabFM model, which has 1.6 billion parameters — and with its "Thinking" mode switched on, Nori apparently does even better, despite running at roughly 2% of TabFM's size.

What makes this interesting isn't just the benchmark bragging rights. It's the workflow problem Synthefy claims to be solving. Fraud detection and dynamic pricing already run on machine learning frameworks like LightGBM and XGBoost, but CEO Somi Agarwal says most enterprises burn weeks preparing data and tuning models for each new problem, and none of that effort carries forward. Point Nori at a fresh table, he argues, and it can produce a solid prediction without a fresh training cycle — turning a weeks-long evaluation into something that takes minutes, because the model already learned from millions of synthetic datasets before it ever saw the customer's data.

The business model looks familiar from the LLM playbook: give away the open foundation layer, then sell the enterprise trimmings. Agarwal is eyeing managed API access, private deployments, security and governance tooling, and large-scale production infrastructure as the actual revenue engine. Early traction backs up the interest — Nori has racked up more than 600,000 downloads in just a few weeks. Wing's Gaurav Garg is betting that structured data models become the next real expansion of the AI market, calling out Synthefy's mix of performance, efficiency and open models as a shot at defining the category before anyone else claims it.

What happens next depends on whether enterprises actually swap out battle-tested tools like XGBoost for a foundation-model approach they haven't lived with for years. Synthefy says the funding goes toward research, hiring engineers and building the next Nori, along with chasing industry partnerships. The download numbers suggest curiosity is already there — the harder test is whether that curiosity turns into production deployments handling real fraud and pricing decisions.