ModelsImplemented and planned capabilities

Implemented and planned capabilities

See which models and tasks Welt runs today, the limits each one is qualified for, and what is still planned.

The bounded upstream TabICLv2, Kumo Tabular Medium, TabDPT 1.3 and Mitra-v2 previews each have separately qualified classification and regression tasks. Arrow / ArrowFM, CDFM and AVICI each qualify observational discovery.

Each model's context shapes, prediction rows per call and classes are in Available models.

The catalogue also reports matching-worker readiness for each task. Other selected families require their own gates; no classical or older-generation fallback is permitted.

Mitra-v2 uses the explicit zero-shot recipe: prediction does not fine-tune weights. This is not a claim to reproduce publisher benchmark recipes.

Available models#

Public catalogue snapshot from 2026-10-15. Context is the training table the model conditions on, or the whole table for causal discovery. It fits when its rows and columns are both within one listed shape, so wider tables take fewer rows. Supervised columns count features without the target; causal discovery counts every variable. Causal discovery needs at least 2 rows and 2 variables. Batch prediction handles larger prediction sets. The live catalogue is authoritative: Client().models() reports current limits and readiness.

What "context" means depends on the task:

• Classification and regression: the training table you pass to fit, counted as rows × feature columns without the target. The model conditions on it when it predicts. Prediction rows are separate: up to 100 per direct call, more with batch prediction.
• Causal discovery: the whole table, with every column a variable.
• Forecasting: the history of one series. Forecasting uses the hidden chronos-2 model, which your workspace must be granted access to; see forecast a time series for its history and horizon limits.

A shape caps both axes at once. TabICLv2 refuses 6,000 rows × 30 features (plus the target): 8,000 × 20 is too narrow and 4,000 × 40 too short. Use 4,000 rows or 20 features instead.

Behavior boundaries#

Upload and batch resources#

Both Client and AsyncClient have these methods:

• Workspace: workspace() and pricing().
• Uploads: create_upload(format=..., size_bytes=..., sha256=..., target=None, schema=None), upload_session(id), put_upload_chunk(id, index, bytes), complete_upload(id) and upload_file(path, target=None, schema=None, format=None, upload_id=None).
• Batch prediction: submit_batch_predict(predictor_id, dataset_id, probabilities=False, idempotency_key=None), job_result(job_id) and batch_payload(job_id, result_reference=None).

On AsyncClient you await the equivalents, including upload, file and batch submission. Job.result() and AsyncJob.result() return batch manifests with explicit partial row ranges. batch_payload() retrieves authenticated JSON until it expires.

File reads and hashing in the async helper are local blocking I/O. The low-level chunk methods let you schedule disk I/O independently.

See the agent CLI reference.

Compatibility with ergodic#

ergodic identifies and estimates effects on a Welt graph, locally. This is the same table as the ergodic compatibility page. The link between the packages is a versioned graph format, not a package version.

• pip install "welt-client[ergodic]" adds ergodic>=0.2,<0.3.
• pip install "ergodic[welt]" adds welt-client>=0.13 on Python 3.11 to 3.13.

Only a graph that Welt declares a native DAG converts:

to_ergodic() in this SDK never decodes: it refuses cdfm and avici results. A decoded graph is the decoder's choice, not the model's claim. See Discover in Welt, estimate in ergodic.

Next steps#