ephemerisTimestral 209M API

Mistral AI · hosted on Ephemeris

Timestral 209M API

Mistral AI's patch transformer with time and variate attention, forecasting 21 quantiles with past-only and known-future covariates. The strongest single member at medium and long horizons in our calibration.

Explicit mode"model": "timestral-209m"
API nametimestral-209m
PublisherMistral AI
Weightsmistralai/Timestral-209M
Size209M
LicenceModified MIT
MultivariateYes
CovariatesPast and known-future
Context read8,192 most recent points

Sources: code · weights

01

REST

Call Timestral 209M directly

cURLUTF-8
curl --request POST \
  --url https://ephemeris.cascade.industries/api/v1/forecast \
  --header "Authorization: Bearer pc_live_your_key" \
  --header "Content-Type: application/json" \
  --data '{
    "mode": "explicit",
    "model": "timestral-209m",
    "series": [{ "values": [112, 118, 132, 129, 121, 135, 148, 148], "freq": "M" }],
    "horizon": 12,
    "quantiles": [0.1, 0.5, 0.9]
  }'
02

Agents

From Claude, Cursor or any MCP client

Add the Ephemeris MCP server, then ask your agent to forecast with timestral-209m. The forecast tool takes mode: "explicit", model: "timestral-209m".

03

Accuracy

How it scores

In our own test on the Forge evaluation pool (38 daily snapshots, scored after the weights were set), Timestral 209M alone was the most accurate member at 256- and 768-step horizons and second to Toto 2 at 64 steps. Adding it to the ensemble improved CRPS by 0.85% against the previous ensemble. Full benchmark results.

No GPU to run

Forecast with Timestral 209M in one request.