Run your first browser forecast
Generate a real model forecast locally without an account, API key, or backend GPU.
Load a target series
Open the playground and choose a template or upload a CSV. Use one numerical target at a consistent cadence. Preserve the observation order, units, and missingness. The browser exports forecast one series at a time; other columns do not become model inputs.
Generate a forecast
Choose a horizon of 1–64 steps and select Generate. The first run downloads a forecast graph (34.8 MB for Tiny or 85.1 MB for Mini) from Hugging Face plus the browser runtime. Download progress appears in the side panel. The worker verifies the model checksum and runs it on your device through WASM; no forecast request goes to a TSFM backend.
Inspect and retain the result
Review the median forecast and P10/P90 bands beside the input history. Drag the forecast origin on the chart into the past to hold out recent data; the playground re-runs the forecast and scores it against what happened. Export the results if you need to retain them. Cancel terminates the worker; removing the downloaded models deletes the cached weights, so the next run downloads them again.