A directory for time series foundation models

TSFM.ai helps you understand the model landscape, compare research evidence, and find an execution option for your forecasting task.

Why the directory exists

Model families differ in architecture, training data, input shapes, task support, and licensing. Published benchmarks answer different questions, and a checkpoint available for download may not be offered by an inference provider. The directory brings model cards, research, comparisons, and execution choices together while keeping those distinctions visible.

Our browser playground runs validated forecast graphs on your device. TSFM does not operate backend GPUs or a hosted model API. Other execution options belong to their providers: follow model links to Hugging Face and upstream implementations to check current availability and deployment requirements.

What we verify

Browser support requires a complete forecast graph, compatible preprocessing and output decoding, and numerical comparison with the original implementation. An encoder export, proxy model, or statistical fallback is not a verified execution of the named model.

Research results remain evidence for the workload they evaluated. We encourage realistic backtests, explicit model revisions, appropriate baselines, and uncertainty evaluation before relying on a forecast for a consequential decision.