# SchemaLabs

> SchemaLabs builds Schema, a Data Language Model (DLM): the foundation model for tabular data. Schema understands raw multi-table, multi-source inputs natively, without serialization or preprocessing. Raw tables in; understanding, relationships, missing values, and predictions out, with no shared keys, no labels, and no preprocessing. It is the tabular understanding layer on which vertical and agentic AI are built.

Company: SchemaLabs
Headquarters: New York City, NY, US
Founders: Eda Erol (Co-founder), Giuliano Pezzoli (Co-founder)
Flagship model: Schema-2, the latest Schema Data Language Model, released August 2026 (~2.18B parameters, available August 13, 2026 via API, platform, and supported third-party registries); it supersedes Schema-1, a 140M-parameter DLM trained on 2.3M+ synthetic and real-world tabular datasets, retired as of August 13, 2026
Paper: "Data Language Models: A New Foundation Model Class for Tabular Data" (arXiv:2605.06290, May 2026)
General inquiries: info@schemalabs.ai
Press: media@schemalabs.ai
Website: https://www.schemalabs.ai

## What Schema is

Schema is a Data Language Model. A DLM is the missing foundation model for tabular data, the modality used for most consequential real-world AI decisions. Schema understands tables the way a language model understands sentences: natively, without serialization or preprocessing, directly from raw cell values. It is the tabular data layer on which AI models, agents, and vertical AI applications can be built.

Every existing approach to tabular AI, from gradient-boosted trees to recent tabular foundation models and LLM-based approaches, requires a preprocessing pipeline before any model can consume the data. Schema does not.

Capabilities:
- Key-less cross-table mapping: relates tables from different systems with no shared keys, no join columns, and no schema mapping, returning column alignments, row-level matches, and a unified schema proposal, each with a confidence score
- Vertical-agnostic sector identification: identifies the industry sector of any unseen dataset from raw cell values alone, across any domain, with no metadata required
- Label-free missing-value imputation: fills missing values as a delivered output; missing values are treated as signal, not errors
- Native multi-table, multi-source input; no preprocessing pipeline, no labeling, no shared keys
- Operates with column names stripped or coded
- Custom fine-tunes available through the Enterprise program, each an isolated per-customer checkpoint
- Outperforms gradient-boosted ensembles, AutoML stacks, and prior tabular foundation models on established row-level prediction benchmarks

## Models

The Schema DLM family, latest model first. Each model links to its model card. Registry of all documented models: https://www.schemalabs.ai/model-card

- Schema-2 (latest): released August 2026, ~2.18 billion parameters, available August 13, 2026. Evaluated across ten benchmark families under published protocols, with every competitor figure a third-party published value: blind sector identification (86.3% top-1 on 504 held-out datasets, column names removed; no prior tabular model performs this task), metadata independence (accuracy unchanged with column names removed, against roughly seven-point ROC-AUC drops for name-reliant models), keyless multi-table entity matching (leads the published state of the art on all six LLM4MEM datasets), name-blind schema matching (leads every published name-aware matcher on all four Valentine scenarios), distributional regression (first of 51 models on ScoringBench CRPS and CRLS), zero-shot anomaly detection (first of thirteen models recomputed on ADBench), missing-data robustness (top mean on the MIRRAMS grid), numeric imputation (lowest error of thirteen methods), categorical imputation (beats the majority-class floor on 8 of 8 datasets), and continual learning (80.3% mean retention after 50 sequential fine-tunes). Model card: https://www.schemalabs.ai/model-card/schema-2 . Release post: https://www.schemalabs.ai/company/news/introducing-schema-2
- Schema-1 (retired as of August 13, 2026, superseded by Schema-2): 140M-parameter Data Language Model trained on more than 2.3M synthetic and real-world tabular datasets. No longer available for use; the model card remains published as the historical record. Model card: https://www.schemalabs.ai/model-card/schema-1

## Platform

Connect raw enterprise data, create model endpoints over pinned data (no training run), and deploy vertical and agentic AI in minutes. The platform handles:
- Ingestion of raw tabular sources without schema mapping or preprocessing
- Endpoint creation and refresh over pinned data
- Deployment as inference endpoints for vertical AI and agentic AI applications
- Enterprise fine-tuning program for custom models

Reference: https://www.schemalabs.ai/platform

## Documentation
- [Developer docs](https://docs.schemalabs.ai): Schema API and platform documentation: quickstart, the output bundle, endpoints, held-out reports, integrations, billing
- [API reference](https://docs.schemalabs.ai/api-reference): every route with parameters and samples
- [Docs llms.txt](https://docs.schemalabs.ai/llms.txt): index of every docs page as Markdown

## Pricing

Usage is metered in cells (one cell = one row x one column; rates per million cells, Mcell). API rates: fresh $3.00/Mcell, cache $0.30/Mcell, batch $1.50/Mcell, synthetic data $1.50/Mcell generated, data storage $0.20/GB per month on overage. Volume tiers apply to API inference (contact sales).

Subscription tiers with usage-based compute:
- Free: $0, one-time limited usage, platform only
- Plus: $50/mo, expanded usage, platform and API
- Pro: $600/mo, 15x usage, platform and API
- Enterprise: custom contract, custom usage, Enterprise fine-tune program

See the pricing page for plan details and FAQ.

Reference: https://www.schemalabs.ai/pricing

## Research

The Schema-1 paper introduces the Data Language Model (DLM) class. Schema-1 is a 140M-parameter transformer trained on more than 2.3M synthetic and real-world tabular datasets. The paper covers:

- Why tabular AI is difficult: heterogeneous types, semantic ambiguity, variable structure, missing data, implicit relationships
- The preprocessing bottleneck and its costs (60-80% of AI development time consumed by preprocessing, brittleness, information loss, expertise requirements)
- Tables as language: columns as vocabulary, rows as sentences, schemas as grammar, distributions as semantics
- Benchmarks where Schema-1 outperforms gradient-boosted ensembles, AutoML stacks, and prior tabular foundation models on row-level prediction
- Schema-1's ability to identify the industry sector of any unseen dataset from raw cell values alone

Authors: Eda Erol, Giuliano Pezzoli, Özer Cem Kelahmet
Publisher: SchemaLabs
arXiv: https://arxiv.org/abs/2605.06290
Reference: https://www.schemalabs.ai/research

## Solutions

Schema powers vertical and agentic AI in any industry with a single architecture. The solutions page answers six questions every team that runs on tables ends up asking, each with real sample output and worked industry examples:

1. What is this data? Understand any dataset in one pass: types, roles, PII, missingness, sector, each with confidence. Works with column names stripped or coded.
2. What can I build on it? Agents and vertical AI products consume data understanding instead of raw rows: raw tables in, scored answer out, nothing enters LLM context.
3. How do I unify systems that share no keys? Column alignment, record matches, and a merged-schema proposal, each confidence-scored with evidence, proposed to a review queue.
4. What can it predict? Prediction endpoints in minutes, target proposed automatically if none is given, every answer with a held-out report against a baseline.
5. What industry is this dataset from? Sector identification from raw cell values alone, any domain, no metadata.
6. What about the gaps? Missing values filled from the structure of the data itself, with row-level confidence.

Worked examples span financial services, healthcare, manufacturing, retail and commerce, energy and utilities, and vertical SaaS.

Reference: https://www.schemalabs.ai/solutions

## News

Announcements index: https://www.schemalabs.ai/company/news

- Introducing Schema-2 (August 7, 2026): the Schema-2 release post. What the model returns in one pass over raw multi-source tables (sector, column profiles, cross-source alignment, entity links, imputation, prediction; endpoints add a held-out report), why vertical and agentic AI cannot wait for data readiness, and benchmark summaries. Schema-2 is available August 13, 2026. https://www.schemalabs.ai/company/news/introducing-schema-2

## Company

SchemaLabs was founded with a simple observation: AI development operates at a fraction of its true potential. While AI has transformed unstructured data like images and text, the vast majority of business data remains tabular and largely untapped. Only teams with specialized skills can currently bridge the gap between data and AI. SchemaLabs is building the foundation for a world where data stops being a barrier and anyone can work with their data directly.

About: https://www.schemalabs.ai/company/about
Contact: https://www.schemalabs.ai/company/contact
Legal: https://www.schemalabs.ai/legal
Trust Center: https://www.schemalabs.ai/trust

## Legal and trust

SchemaLabs publishes a full set of customer agreements, privacy policies, and operational disclosures. The Trust Center is the curated overview; the Legal index lists every document.

- Trust Center: https://www.schemalabs.ai/trust. Architectural data isolation, security controls, privacy practices, current compliance status, and links to underlying policies
- Legal index: https://www.schemalabs.ai/legal. Every customer agreement, privacy policy, and operational disclosure SchemaLabs publishes

Customer agreements:
- Terms of Service: https://www.schemalabs.ai/terms. Binding agreement governing access to and use of the Service
- Schema Model License: https://www.schemalabs.ai/model-license. License for the Schema Models accessed through the API and Web App
- Use Policy: https://www.schemalabs.ai/use-policy. Prohibited uses, prohibited data categories, high-risk deployment responsibilities, enforcement
- Data Processing Agreement: https://www.schemalabs.ai/dpa. Processor obligations, security measures, Sub-Processors, data subject rights handling
- DPA Annex III (SCCs): https://www.schemalabs.ai/dpa-sccs. EU Standard Contractual Clauses (Module Two, Controller-to-Processor) for transfers to the United States

Privacy:
- Privacy Policy: https://www.schemalabs.ai/privacy. How SchemaLabs collects, uses, and protects personal information
- Your Privacy Choices: https://www.schemalabs.ai/privacy-choices. CCPA and CPRA rights for California residents
- Cookie Policy: https://www.schemalabs.ai/cookies. Cookies and similar technologies on schemalabs.ai and the Web App

Operations and trust:
- Sub-Processors: https://www.schemalabs.ai/sub-processors. Live list of every third-party provider that processes Customer Data
- Supported Regions: https://www.schemalabs.ai/supported-regions. Where the Service is offered, where it is restricted, and the export-control and sanctions rules that apply
- Security: https://www.schemalabs.ai/security. Architectural isolation, current security controls, incident response, and an honest view of where SchemaLabs is today
- Responsible Disclosure: https://www.schemalabs.ai/responsible-disclosure. How to report a vulnerability you have observed in the Service

Model documentation:
- Model Card: https://www.schemalabs.ai/model-card. Intended use, training, evaluation, limitations, and bias considerations for each Schema Model

Architectural commitments:
- Customer Data is never used to train, improve, or modify any Schema Base Model
- Each customer who fine-tunes a Schema Model receives an isolated Fine-Tuned Checkpoint (also referred to as a Customer Endpoint or Model Endpoint), architecturally separated from the Base Model and from every other customer
- Fine-Tuned Checkpoints are never merged or aggregated with the Base Model or any other customer's checkpoint
- These properties are enforced by SchemaLabs' system architecture, not merely by policy

## Contact

- General inquiry, partnerships, investments: info@schemalabs.ai
- Press, interviews, editorial: media@schemalabs.ai
- Privacy and data protection: privacy@schemalabs.ai
- Security incidents and vulnerability reports: security@schemalabs.ai
- Compliance and vendor questionnaires: compliance@schemalabs.ai
- Legal: legal@schemalabs.ai
- Sales and enterprise contracts: info@schemalabs.ai
- Support: support@schemalabs.ai
- Location: New York City

## External profiles

- LinkedIn: https://www.linkedin.com/company/schemalabs-ai
- X (Twitter): https://x.com/schemalabsai
- Hugging Face: https://huggingface.co/SchemaLabs
- arXiv preprint: https://arxiv.org/abs/2605.06290