Data Language Models: A New Foundation Model Class for Tabular Data

Researchers introduce Schema-1, the first Data Language Model (DLM) designed to natively understand tabular data without preprocessing, similar to how language models understand text. The 140M-parameter model trained on 2.3M datasets outperforms gradient-boosted trees, AutoML systems, and existing tabular foundation models on prediction benchmarks and demonstrates superior performance on missing value imputation and dataset classification tasks.

The introduction of Schema-1 represents a significant maturation in AI's ability to handle tabular data, which powers most real-world business applications yet has remained the least developed modality in the foundation model landscape. While text, images, and audio each have native foundation models that eliminate preprocessing friction, tabular AI has required custom data pipelines before any model can operate. This architectural gap has limited the efficiency and capability of data-driven systems across finance, healthcare, logistics, and enterprise software.

Schema-1 addresses this gap by treating tabular data as a language unto itself, processing raw cell values directly without serialization or feature engineering. The model's 140M parameters were trained on both synthetic and real-world datasets, enabling it to understand structural patterns in tabular formats. Its performance advantage over gradient-boosted ensembles and AutoML stacks is meaningful because these methods currently dominate enterprise AI deployments. The missing value reconstruction capability is particularly noteworthy: the model outperforms both classical statistical imputation and large language models, suggesting that learned distributional geometry of actual data proves more valuable than general world knowledge for this task.

For the AI industry, this development accelerates the abstraction layer between raw data and intelligent applications. Enterprise software vendors, analytics platforms, and AI infrastructure providers will likely integrate or build upon tabular foundation models as standard components. The ability to classify datasets by industry from raw values alone indicates Schema-1 captures domain-specific patterns without explicit training signals, opening possibilities for zero-shot or few-shot applications across verticals.

Future developments may include multimodal versions combining tabular, text, and structured data understanding, as well as deployment optimizations for production systems handling billions of rows.

• →Schema-1 is the first foundation model designed to natively understand tabular data without preprocessing or serialization, matching capabilities of language models for text.
• →The model outperforms gradient-boosted trees, AutoML stacks, and existing tabular foundation models on standard benchmarks and missing value imputation tasks.
• →Schema-1 can identify industry sectors from raw cell values alone, a capability no prior tabular model demonstrated, suggesting learned structural understanding of domain-specific patterns.
• →Tabular data powers most real-world AI decisions but historically lacked native foundation model support, creating preprocessing bottlenecks throughout the AI stack.
• →This development enables direct integration of raw tabular data into AI systems, eliminating custom feature engineering pipelines that currently stand between data sources and applications.