Google TimesFM — hosted and ready
Run TimesFM 2.0 through the TSFM.ai API. A patched-decoder architecture trained on Google Trends and public time series corpora, with strong zero-shot performance on long-horizon forecasts.
{
"model": "google/timesfm-2.0",
"inputs": [{
"item_id": "energy-load",
"target": [320, 340, 310, 355, 370, 345, 360],
"start": "2026-03-01T00:00:00Z"
}],
"parameters": {
"prediction_length": 30,
"freq": "D",
"quantiles": [0.1, 0.5, 0.9]
}
}Same request shape as every other model on TSFM.ai.
Architecture
200M-parameter patched-decoder transformer
Context
Up to 2048 time steps of history
Strength
Strong long-horizon and trend-following performance
Why teams choose TimesFM
Patched-decoder design
TimesFM processes input time series in patches rather than individual tokens, improving computational efficiency and allowing the model to capture multi-scale patterns in a single forward pass.
Strong on long horizons
TimesFM was trained on data distributions that emphasize longer-range dependencies. This makes it particularly effective for 14-day, 30-day, and multi-month forecast horizons.
Single well-tuned model
Unlike model families with multiple size variants, TimesFM ships as one 200M-parameter model that performs consistently well across domains — no size selection complexity.
Compare with ChronosGetting started with TimesFM
- 1
Get an API key
Sign up for a free account and generate a key from your dashboard. TimesFM is available on every plan.
- 2
Send a forecast request
POST to /v1/forecast with model set to google/timesfm-2.0. Pass your time series and desired prediction length.
- 3
Compare with other models
Run the same request with amazon/chronos-bolt-base or salesforce/moirai-1.1-r-base. The API shape is identical — only the model ID changes.
TimesFM at a glance
Frequently Asked Questions
Try TimesFM on your data
Send a series, pick a horizon, and see how TimesFM 2.0 performs on your forecasting workload. No setup required.