Tabular foundation model
Tabular foundation model
Documentation for the tabular foundation model: tables and databases in, data understanding, relationships, missing values, and predictions out. No shared keys, no labels, no preprocessing.
Informational documentation. This section explains how a tabular foundation model of the Data Language Model class works: what goes in, what comes out, and how results are evaluated. It is based on the paper Data Language Models: A New Foundation Model Class for Tabular Data (arXiv:2605.06290). It is reference material, not a hosted service: hosts such as
api.example.com, keys, and ids are placeholders.Background
Background
What it is
What a tabular foundation model is, what it reads, what it returns, and how it differs from training a model per table.
Read Background
Research
The paper behind the model: why tabular data has no native foundation model, and what a Data Language Model is.
Read Background
Solutions
One foundation model across every industry. Your data ships as tables: exports, extracts, feeds, logs. The model reads them raw and answers six questions that every team running on tables ends up asking.
Read The model
The model
Tabular foundation models
What a tabular foundation model is, the landscape of approaches from boosted trees to data language models, and why reading the raw table matters.
Read The model
Architecture
How a tabular foundation model is built: cell encoding, axial attention, Perceiver-style global reasoning, task heads, adapters, and one merged checkpoint.
Read The model
Cell encoding
How a tabular foundation model turns each raw cell into a vector: value, type, position and explicit missingness, with column identity zeroed.
Read The model
Pre-training
How a tabular foundation model learns before it sees your table: masked cell modeling, synthetic data priors, distributed training, and checks.
Read The model
In-context prediction
How a tabular foundation model predicts without training on your table: labelled rows act as a support set, and new rows are read off by comparison.
Read The model
Continual learning for tables: learning new tasks without forgetting old ones
How a tabular foundation model learns task after task without forgetting: adapters, replay of 48 raw rows per task, what failed, and what stays open.
Read The model
Column name independence: how a tabular model works without headers
Read The model
Compression and serving: running a 2.18 billion parameter tabular model
How a tabular foundation model is compressed to half, 8-bit and 4-bit precision, how fidelity is checked, and what changes when it serves requests.
Read The model
Capabilities
What a tabular foundation model returns in one pass: sector, column profile, cross-table map, imputation, predictions, and a held-out score.
Read The model
Evaluation
How a tabular foundation model is evaluated: held-out splits, benchmark protocols, sealed configurations, contamination checks, and what a score claims.
Read The model
Limitations
What a tabular foundation model does not do: data limits, task limits, benchmark caveats and open problems, so results are read correctly.
Read Benchmark stories
Benchmark stories
Benchmarks
What each benchmark measures, how I built the capability that answers it, and what made the difference, in the order the work was done.
Read Benchmark stories
Row-level prediction benchmark: how the tabular foundation model was built and measured
How a tabular foundation model was built and measured on row-level prediction boards: classification and regression, ranked by Elo against tuned pipelines.
Read Benchmark stories
Continual learning benchmark: keeping old tasks while learning new ones
How a tabular foundation model learns task after task without forgetting: the failures, the replay recipe that stayed, and what stayed open.
Read Benchmark stories
Anomaly detection benchmark: finding rows that do not fit, with no labels
How a tabular foundation model ranks anomalous rows in the semi-supervised ADBench setting, and how label-free channel selection made the difference.
Read Benchmark stories
Probabilistic regression benchmark: predicting a whole distribution, scored by CRPS and CRLS
How a tabular foundation model predicts full distributions, not single numbers, on ScoringBench, and how calibration fixed the CRLS gap.
Read Benchmark stories
Metadata independence benchmark: predicting with no column names
How a tabular foundation model keeps its accuracy when column names are removed or scrambled, and how I measured that fairly against rivals.
Read Benchmark stories
Blind sector identification: naming a dataset's industry from raw cells
How a tabular foundation model names the industry of an unseen table from cell values alone, with no headers, labels or metadata, and how I tested it.
Read Benchmark stories
Missing-value imputation benchmark: filling the gaps in real tables
How a tabular foundation model fills missing numeric and categorical cells, and how a hunt for fifteen measurement errors produced a fair first place.
Read Benchmark stories
Entity and schema matching benchmark: joining tables that share no keys
How a tabular foundation model matches records and columns across tables with no keys and no column names, and how I measured it against name-aware rivals.
Read Benchmark stories
Missing-data robustness benchmark: predicting when the features have holes
How a tabular foundation model keeps its predictions when values are missing from the features, and why I checked the ruler before trusting any number.
Read Benchmark stories
One checkpoint, then compression: packaging a whole benchmark result
How the core, every head and every adapter were merged into one checkpoint of about 2.18 billion parameters, then compressed and audited before release.
Read Using the model
Using the model
Quickstart
Get an API key, quick-run the model across sources that share no key, and publish the same configuration as an endpoint, in about five minutes.
Read Using the model
Introduction
What the model is, what one call returns, and the two ways to consume it.
Read Using the model
Authentication
Read Using the model
Models
The foundation models you can run on and create endpoints on, how to choose a base, and how an upgrade moves an endpoint to a newer base.
Read How it works
How it works
Data
Read How it works
Outputs
The one response every pass returns: sector, column profile, cross-table map, imputation, predictions, and on endpoints the held-out block.
Read How it works
Endpoints
Create an endpoint over pinned data, serve it with a serve-scoped key, refresh when the source changes, and upgrade when a new base ships.
Read How it works
Held-out reports
How the held-out score is computed by in-context split evaluation, what every report contains, and exactly what the number claims.
Read Integrations
Integrations
Agents and LLMs
Give your agent the model as a tool, over MCP from any framework or over REST with one function-calling definition, and narrate the bundle with your own LLM.
Read Integrations
MCP server
Use the model from your AI. One remote server for Claude, ChatGPT, Cursor, VS Code, Gemini CLI, and any MCP client. Sign in or use an API key. Everything on the platform and the API, with the same output bundle.
Read Integrations
Workflows and data stack
Call the model from any HTTP node, schedule syncs and refreshes from your orchestrator, and land bundles in the warehouse.
Read Resources
Resources
Troubleshooting
Every response the API returns when something needs your attention, what it is telling you, and the concrete next step.
Read Resources
Glossary
The product vocabulary, one entry per term, identical across the API and the platform.
Read Key facts
What is a large tabular foundation model?
A large transformer pre-trained on a very wide range of tables, so it can predict, fill in missing values, flag anomalies, and match records on a new table in one pass, without training a new model for each dataset.
What is a Data Language Model (DLM)?
A Data Language Model is a foundation model that reads raw tables directly, the way a language model reads text, with no serialization, no preprocessing, and no column names. The category is defined in the arXiv paper 2605.06290 by Özer Cem KELAHMET.
What is Schema-2?
Schema-2 is the large tabular foundation model built by Özer Cem KELAHMET. It has more than 1.5 billion parameters, uses axial attention over rows and columns, a Perceiver-style global path, and a schema path, and one checkpoint of about 2.18 billion parameters carries every capability.
What can the model do?
Row-level prediction, missing-value imputation, anomaly detection, probabilistic regression, continual learning across tasks, entity and schema matching, blind sector identification, and prediction that does not depend on column names.
Who is Özer Cem KELAHMET?
A technology and AI leader who builds Data Language Models, large tabular foundation models, and real-time AI systems, end to end from research and architecture to production.
Where can I read the paper?
The paper is on arXiv: https://arxiv.org/abs/2605.06290. The documentation on this site covers the architecture, pre-training, evaluation, and every benchmark.