AI firms push beyond LLMs with new model categories

AI developers are increasingly releasing models that differ from large language models, with recent launches spanning tabular foundation models, decision-only models and code execution models, according to a roundup by Digital Today.

Nvidia recently published Kumo Tabular, an open foundation model for predicting structured, table-form data such as sales records, transaction histories and customer information, on the Hugging Face platform. The model appears to be based on technology from Kumo AI, an enterprise predictive AI software startup Nvidia acquired in June.

Others have moved on decision-focused models. TypeSafeAI, founded by OpenAI alumnus Diogo Almeida, unveiled Jev in September; given a complex situation, a narrow question and a list of possible answers, Jev returns decisions and probabilities instead of sentences. Microsoft released Microsoft-Decision-1, a model specialized for fast decision-making, and Cloudflare published a preview of Claretic, linked to its decision model Cleft, which extracts names, dates and amounts from unstructured material such as emails and receipt photos and estimates values not present in the material.

Nums AI released its own LFM, Causilo, and has begun targeting the market. Israel-based DoubleAI, backed by Nvidia, is promoting CEM as a new AI category; its first product, Maxwell, is said to improve GPU and CPU performance to address rising computing costs. DoubleAI charges customers a share of the savings it generates, and one early customer cut computing costs by 67%, according to the company. LTX doubled model inference speed with Maxwell, according to the Wall Street Journal.

In background, DoubleAI has raised $220 million to date from Lightspeed Venture Partners, Bessemer Venture Partners, Nvidia and Dell, and SAP acquired Prior Labs, a company specializing in tabular foundation models. Gartner named Large Quantitative Models (LQM) as one of the keywords to watch in the tech market in 2027.