tabSutra 39M v1.0
tabSutra is a foundation model for structured data: time series and, in later versions, tables and relational
databases. This card describes the first released checkpoint, tabSutra_39M_v1.0, evaluated on GIFT-Eval and
fev-bench as a zero-shot forecaster.
The weights are not public. The model will be available through an API and for on-premise deployment; see tabsutra.com.
At a glance
| Parameters | 39.1M (29.2M used by a forecast; the rest serve tabular tasks and were not trained in this checkpoint) |
| History | up to 9,984 past steps, read in 1,024 rows |
| Horizon | up to 1,024 steps ahead |
| Output | 19 quantiles, 0.05 to 0.95 |
| Inputs | the target; known-ahead, past-only and static covariates, numeric and categorical; missing values; irregular timestamps; several related series in one call |
| Training data | synthetic only: about 3.6 million generated tables; no real-world series, no benchmark data |
| Hardware | one NVIDIA RTX PRO 6000 Blackwell Max-Q (96 GB) |
| Training | 280,000 steps, about 9 days |
| Inference precision | bfloat16 |
What the model is
The model reads a table of rows and cells. A time series is a table whose rows are time steps. A panel of related series is one table with several entities. A plain table has rows with no clock. A table whose rows are not independent is the same table with its ordering kept: rows with a timestamp (transactions, events, repeated measurements of an entity), rows in a defined sequence, or rows in a given order with no time at all, where the position serves as the clock. The same forward pass serves all of them, and the forecasting head, the classification head and the regression head sit on the same row states.
Properties that follow from the design:
- No column names. Columns are described by their content, never by a name or a position. Permuting the columns of a table gives the same prediction.
- Known-time rule. Every row carries the time its features became known and the time its label became known. A row may use another row's features only if they were known at or before its own time, and another row's label only if it was known strictly before. This is what makes forecasting, backtesting and label leakage well defined in one mechanism, for regular and irregular timestamps alike.
- Multi-resolution history. A series of up to 9,984 past steps is read in 1,024 rows: the oldest at 64-step resolution, then 8-step rows, then 768 rows at single-step resolution. Horizons up to 1,024 steps ahead.
- Covariates. Known-ahead numeric and categorical columns, past-only columns, static numeric and categorical attributes, missing values, and known-ahead columns that are known for only part of the horizon.
- Entities. Related series (the stores of a chain, the sensors of a site) are the entities of one table and read each other's history with as-of alignment; up to 1,024 entities in one call. Multivariate series are handled as entities or as columns.
- Calendar. Timestamps give the model calendar phases (hour, day, week, month, year) at any sampling rate from seconds to years.
Output
19 quantiles (0.05 to 0.95 in steps of 0.05) for every forecast step, in the series' own units. Point forecasts use the median.
Training data
This checkpoint was trained on synthetic data only. No real-world time series was used. No GIFT-Eval dataset, train or test split, and no dataset of fev-bench, TIME or BOOM was part of the training data in any form.
Training ran for 280,000 steps over about 3.6 million generated tables, on one NVIDIA RTX PRO 6000 Blackwell Max-Q (96 GB), in about 9 days.
Inference-time post-processing
Two steps are applied to the raw quantiles and are part of the evaluated model:
- Nonnegative: when a series' history has no negative value, every quantile is floored at zero.
- Integer rounding: when a series' history is all whole numbers, the median is rounded to the nearest whole number.
Both are decided per series from its own history. No test-time ensembling, no averaging. Without these two steps the same checkpoint scores CRPS 0.534 and MASE 0.781 on GIFT-Eval, and 42.1% / 29.7% on fev-bench.
Evaluation
GIFT-Eval:
| CRPS (MWSQL) | MASE | |
|---|---|---|
| tabSutra_39M_v1.0 | 0.532 | 0.778 |
Submission type: zero-shot; test-data leakage: no.
GIFT-Eval-CTX (10,000 instances from held-out GIFT-Eval test regions, 78 dataset–term groups; history-only; scored as: per-group mean MASE relative to seasonal naive, geometric mean over groups, lower is better):
| short | medium | long | overall | |
|---|---|---|---|---|
| tabSutra_39M_v1.0 | 0.745 | 0.821 | 0.856 | 0.792 |
Only each instance's history, horizon and frequency were read; the benchmark's structured effects were not used.
fev-bench (skill score is 1 minus the geometric mean of the ratio to seasonal naive, higher is better):
| SQL skill | MASE skill | |
|---|---|---|
| tabSutra_39M_v1.0 | 42.2% | 30.4% |
Same checkpoint, same two post-processing steps as on GIFT-Eval.
Intended use
Demand and sales forecasting at SKU, store and category level; operations and sensor series; any forecasting problem with covariates, hierarchies and new entities, with or without long history. The tabular and relational capabilities of the architecture are trained in later checkpoints.
Limitations
- Trained on synthetic data only.
- Long horizons on near-deterministic daily profiles and very long context still trail the largest published models on GIFT-Eval.
- The tabular heads of this checkpoint are untrained; use it for forecasting only.