TFMs handle numeric data as numbers, not text tokens, making them far more accurate for financial spreadsheets than standard LLMs. Google and Nvidia are leading the development.
A new class of AI models is starting to handle the kind of data that powers most financial analysis: rows and columns of numbers, dates, and categories. Tabular foundation models, or TFMs, are built to do what large language models often do poorly – treat numeric data as numbers, not text tokens.
LLMs like GPT-5 or Claude convert everything into tokens. A customer age of 36 becomes the characters "3" and "6", not a numeric value. That makes it hard for the model to find mathematical relationships between age and purchase amount, or to spot outliers in a column. TFMs sidestep that by encoding numeric fields as numbers from the start. They also handle missing values, out-of-range entries, and mixed data types – the messy reality of any real-world spreadsheet.
The technology is still early. The best-known TFM is TabFM from Google Research, which uses alternating attention across rows and columns. Another is TabPFN from Prior Labs. Nvidia and Kumo have developed KumoRFM. These models are typically trained on synthetic data crafted to resemble real tabular data, rather than on internet text. The idea is to expose the model to the kinds of errors and anomalies it will encounter in practice.
For financial analysts, the implications are direct. Portfolio data, risk reports, transaction logs, and earnings tables all fit the tabular format. A TFM can surface correlations, flag unusual entries, or estimate missing values without the user having to write complex prompts or clean the data first. Current LLMs require heavy prompt engineering to even attempt this, and still produce unreliable results.
Nvidia, which has an Alpha Score of 71, is one of the companies pushing the technology forward through its KumoRFM work. The model is designed for enterprise tabular data, including financial and customer datasets. Google's TabFM paper, released in June 2026, shows zero-shot performance on benchmark tabular tasks that beats tuned LLMs.
Still, TFMs are not plug-and-play. They require careful setup and are best suited for specific domains. A model trained on medical data will not work well on inventory data without retraining. The current workaround is to use a broad TFM or to fine-tune one for a particular sector.
The longer-term direction, as several researchers have noted, is hybrid models. An LLM would serve as the natural language interface, and a TFM would handle the tabular calculations behind the scenes. That would let an analyst ask "show me the accounts with the highest revenue growth last quarter" and get a correct answer without the LLM mangling the numbers.
For now, anyone working with spreadsheets and looking for AI help should know that the right tool exists – but it is not the same tool that writes emails or answers trivia. The gap between what LLMs can do and what TFMs can do for tabular data is wide, and the TFMs are only beginning to close it.
Drafted by a large language model from the source reporting linked above, then screened by automated publishing checks. It is not read by a journalist before publication. Some articles cite our Alpha Score. Verify prices and figures against the original source. Educational coverage, not personalized advice.