Time-series foundation models as a service
One API. 49+ pretrained forecasters. No fine-tuning required.
What we host
Every major open-weights time-series foundation model, served behind one consistent inference API. See the full catalog collection for the exact 49 models you can call today.
Why a dedicated provider
General-purpose LLM inference stacks are a bad fit for forecasting. Time-series models have
narrow context windows, variable history lengths, quantile outputs, exogenous covariates, and
probabilistic sampling — none of which map cleanly onto OpenAI-style APIs. We built TSFM.ai for
this surface: past_values, past_timestamps,
past_covariates, future_covariates, static_covariates,
quantiles, and num_samples are first-class.
Get started
curl -X POST https://api.tsfm.ai/v1/forecast \
-H "Authorization: Bearer $TSFM_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "amazon/chronos-2",
"inputs": [{"target": [10, 12, 11, 13, 14, 15, 14, 16, 18, 17]}],
"parameters": {"prediction_length": 24, "quantiles": [0.1, 0.5, 0.9]}
}'from tsfm import Tsfm
client = Tsfm()
forecast = client.forecast(
model="amazon/chronos-2",
inputs=[{"target": [10, 12, 11, 13, 14, 15, 14, 16, 18, 17]}],
parameters={"prediction_length": 24, "quantiles": [0.1, 0.5, 0.9]},
)
print(forecast.predictions[0].mean)Benchmarks
We publish continuously-updated scores for every hosted model on GIFT-Eval and Impermanent.