Synthefy's $6.5M seed bets numbers need their own foundation models, not a tokenizer

Synthefy Inc. told SiliconANGLE it has raised $6.5 million in seed funding led by Wing Venture Capital, with Haystack, Samsung Next, Canonical Crypto and Lightscape participating alongside angel investors from OpenAI, Microsoft and Meta [s1c1][s1c2]. The dollar figure is unremarkable at seed; the category label attached to it is the part worth tracking, because it stakes out a position on where numerical work should live.

The company calls its models Structured Data Foundation Models, or SDFMs, and says they are trained on numerical data rather than text, on the argument that ingesting large volumes of number-crunching data lets a model preserve the relationships inside time-series data and tables [s1c3][s1c4]. SiliconANGLE frames the pitch as doing for numbers what large language models did for words [s1c1]. That framing implies text models handle tables badly. Note what the reported evidence actually compares against: Google's 1.6-billion-parameter TabFM, and the gradient-boosting frameworks LightGBM and XGBoost that enterprises already use for this work [s1c6][s1c7]. No general-purpose LLM baseline appears in the report [s1c14]. The failure mode being sold against is not tokenization. It is the cost of building a fresh tabular model per problem.

Read the operator complaint on its own terms and it is a familiar one. Co-founder and Chief Executive Somi Agarwal told SiliconANGLE that most enterprises spend weeks on data preparation, training and fine-tuning before a model performs reliably [s1c8]. "That work does not compound," he said. "Each new fraud, pricing or forecasting problem starts again." He says pointing Nori, the company's first open-source SDFM, at a new table can bring initial evaluation down from weeks or months to minutes [s1c9]. Nori was released quietly a few weeks before the funding announcement [s1c5], and Agarwal says it has learned from millions of synthetic datasets, so it arrives at a new problem "with experience rather than starting from zero" [s1c10].

The benchmark claim is the company's own. Synthefy says a 30-million-parameter Nori outperformed TabFM, and does better still with its "Thinking" capability enabled, at 2 percent of the Google model's size [s1c6]. The arithmetic is roughly consistent: 30 million is about 1.9 percent of 1.6 billion, a parameter ratio near 53 to 1 [s1c13]. Nobody outside the company has published a reproduction in this account. Adoption so far is measured in downloads, more than 600,000 for the first Nori version within weeks of release [s1c11], which indicates curiosity rather than production usage.

The business is the usual open-core shape. Access comes through the open model, a managed API, or deployment inside a customer's own environment [s1c12], and Agarwal says the revenue is in the enterprise wrapper: managed API usage, proprietary features, private deployments, security and governance controls, integrations, support and production infrastructure [s1c15]. Wing founding partner Gaurav Garg says he is betting structured data models become the next major expansion of the AI model market [s1c16]. The funding goes to research, engineering hires and the next generation of Nori [s1c17].

Watch for an independent reproduction of the TabFM comparison, and for whether the zero-tuning claim survives contact with real enterprise tables, where leakage and drift usually consume the weeks that were supposedly saved.