News · Model release

Introducing Schema-2

Schema-2 is a Data Language Model: AI that understands business data directly. It takes tables and databases as they are, across any number of sources, and returns what the data is, how the sources relate, what is missing, and what can be predicted from it. No shared keys, no labels, no preprocessing step.

Availability

Schema-2 is available to use starting August 13, 2026, through the SchemaLabs API, platform, and supported third-party registries. Rates are in Pricing.

Every AI initiative built on business data begins the same way. Before a model can run, someone has to establish what the fields mean, how the sources relate to one another, which records can be trusted, and what the model should predict. That work is repeated for every new source and every new project, and it is the step that most often stalls the initiative.

Schema-2 performs it in a single call, on the data as it stands.

In practice

Multiple tables that share no key. One request.

The data stays as it is. The understanding happens inside the model: no mapping, no labels, no preprocessing.

In · raw tables · shared keys: none
crm_contacts 2,300 rows
full_namephoneemailsignup_date
Ana Kaya555-0142ana.k@…2024-11-03
Devraj Nair555-0199d.nair@…2024-12-01
Mira Ostrom555-0210null2025-01-17
cards_db 2,300 rows
holderphone_e164card_brandcredit_limit
KAYA, ANA+1 555 0142Visa12000
NAIR, DEVRAJ+1 555 0199Mastercard6500
OSTROM, MIRA+1 555 0210Amexnull
txn_2025q1 184,902 rows
c1c2c3c4
+1555014234.90581204/01/25
+155502101290.00573209/01/25
+1555019912.40541111/01/25
Out · one pass · POST /v2/create
"endpoint": { "cardbrand", status: live, mode: create }
"feed": { multi_table · 3 tables · 2,300 records }
"sector": { "consumer financial services", 0.89 }
{ "customer relationship mgmt", 0.74 }
"profile": { phone "phone", 0.97, pii: true }
{ email "email", 0.98, missing 2% }
{ c3 "merchant_category_code", 0.88 }
"alignment": { phone phone_e164, 0.96 }
{ full_name holder, 0.90 }
"entities": { 2,180 of 2,300 linked, keys: none }
"unified": { 7 cols · 2 shared · 2 a-only · 3 b-only }
"imputation": { credit_limit 8400.0, 0.81 }
"task": { auto classification · 4 classes }
"target": { auto cards_db.card_brand }
"scored": { 2,208 rows · 92 dropped, no target }
"held_out": { acc 0.91 · auc 0.95 · base 0.61 }
"split": { in_context_holdout · 1766/442 · seed 42 }
"per_class": { Visa .94 · Mastercard .90 · Amex .86 }
"weakest": { "Discover: 0.79" }
"report": { "/v2/reports/r_3c7e" }
Why this matters now

Vertical and agentic AI cannot wait for business data readiness

A vertical AI meets customer's raw data on the first day of onboarding; AI agents meet new sources mid-task, at runtime, with nobody available to explain it. Neither can pause for a quarter while a data team prepares that data, and no human team can sit inside an inference loop.

01 · NEW SOURCES, AT RUNTIME Partner file drop Schema change in production Third-party API response Database from an acquisition New customer onboarding 02 · VERTICAL OR AGENTIC AI Meets data it has never seen 03 · DATA LANGUAGE MODEL Schema-2 04 · DATA UNDERSTANDING What the data is How the sources join What is missing or odd What it can predict returned to the AI · no ticket, no preprocessing, no wait
Benchmarks

Performance, measured

Benchmark summaries below. To review all results in full detail, protocols, datasets, per-dataset tables and competitor sources, go to the Schema-2 model card.

Identifies the data without being told

Given a table it has never seen, with column names removed and no documentation, Schema-2 names the industry sector from cell values alone. Correct on the first attempt in 86.3% of cases.

Schema-2 · top-186.3% Schema-2 · top-591.1% Random guess0.01% 0%50%100% Top-1 and top-5 accuracy, higher is better.

Use it forCataloguing a data lake, routing incoming files to the right pipeline, triaging a new dataset before anyone opens it.

Holds accuracy when column names are gone

Systems that depend on column names lose roughly seven points of accuracy when those names are absent, obscure, or stripped for privacy. Schema-2 does not move.

0.960.930.900.870.84 0.92300.86580.8541 Full namesRandomNo names Schema-2TabuLa-8BConTextTab

Use it forLegacy systems, acquired databases and anonymised exports, with no renaming project or data dictionary needed first.

Matches records with no shared ID

The same customer in two systems that never shared a key. Schema-2 leads the published state of the art on all six standard datasets. On Person, the hardest, it reaches 79.1 against 41.8.

Geo97.2 Music-20K93.8 Music-200K92.3 Music-2M79.8 Person79.1 Shopee40.7 050100 Schema-2, no keysbest published

Use it forCustomer deduplication, post-acquisition record merges and supplier consolidation across systems with no common ID.

Aligns columns across schemas, name-blind

Which column here corresponds to which column there, decided from the values alone. Every published matcher reads the schema names. Schema-2 does not, and still leads all four scenarios.

Joinable0.990.92 Sem.-joinable0.990.94 Unionable1.000.97 View-unionable0.940.91 0.00.51.0 Schema-2, blindBest published, name-aware

Use it forData migrations, warehouse consolidation and partner integrations where the column names do not line up.

Predicts ranges, not single numbers

Point estimate says a claim will cost $1,200; distribution says how likely every other outcome is. Across 102 datasets and 51 models, Schema-2 places first, with a 4.4x margin.

Schema-20.048 nori_30m0.2120 nori0.2453 TabPFN-30.3792 RealTabPFN-2.5 ft0.3857 TabICLv2 ft0.4018 0.00.20.4 Median CRPS across 102 datasets, lower is better.

Use it forPricing, demand planning, reserve setting and risk, where you need the range and not a single number.

Keeps predicting when data goes missing

Training and serving data going incomplete in different ways at the same time, which is what production actually does. Schema-2 holds the top mean across the grid, narrowly.

Schema-20.8943 MIRRAMS0.8878 CatBoost0.8697 SAINT0.8608 XGBoost0.8582 0.00.51.0 Mean ROC-AUC, higher is better.

Use it forProduction scoring where fields drop out, sensors fail or forms arrive half finished.

Beats every published imputation method

Averaged over nine ways of hiding data across twenty real datasets, Schema-2's reconstructions carry 1.8x less error than the next best frontier LLM and 2.3x less than the best classical method.

Schema-20.133 Gemini 3.0 Flash0.235 Claude 4.5 Sonnet0.237 missForest0.302 MICE0.306 TabPFN0.448 0.00.250.5 Mean NRMSE across nine conditions, lower is better.

Use it forRepairing partial records before they reach a report, a downstream model or a regulator.

Fills categorical gaps, not just numbers

A missing diagnosis code or an empty churn flag cannot be averaged in, and filling with the most common value erases the rare categories that carry the signal. Schema-2 beats that floor on all eight datasets.

Mean0.54280.3649 stroke0.81100.4623 hepatitis0.91720.7232 chronic0.33890.1023 0.00.51.0 Schema-2Most-common value

Use it forMissing diagnosis codes, blank churn flags and unset status fields, where the rare category is the one that matters.

Detects anomalies with no labelled examples

Zero-shot: no labelled examples of anomalous rows are required. Schema-2 leads the 57-dataset standard suite, though the top of this field is closely grouped, and the chart shows that.

Schema-20.8404 OutFormer0.8242 DTE-NP0.8149 TabPFN-OD0.8101 kNN0.8086 DTE-C0.8037 FoMo-0D0.7989 0.00.51.0 Mean AUROC across 57 datasets, higher is better.

Use it forFraud screening, equipment monitoring and data-quality checks with no labelled history to train on.

Takes on new data without a full retrain

After fifty fine-tunes on fifty different domains, one after another, Schema-2 keeps 80.3% of what it could do on each. A gradient-boosted stack keeps nothing: it is rebuilt from scratch every time.

Schema-280.3% Sequential baseline51.2% Tree-based stacks0% 0%50%100% Mean retention, higher is better.

Use it forOne endpoint that follows the business into new products, regions and datasets instead of a rebuild per use case.

Go deeper

The model card has the full record

Protocols, per-dataset tables, every competitor source, the four benchmark families this post omits, and the limitations in full.

Read the model card