TimesFM 2.5 · 200MZero-shot · no training

Forecasting as an HTTP call.

Post a time series, get a probabilistic forecast back. VecTime Cloud runs Google's TimesFM foundation model behind a metered API — no model to train, no features to engineer, nothing to deploy.

Start free — 300 callsRead the API docs

No card for the free tierQuantile bands includedLimits in every response header

Daily orders → next 24 steps

zero-shot

HistoryForecast (median)P10–P90 band

endpoint. POST a series, get a forecast.

decile quantiles on every forecast, q10–q90.

512

series per request on the largest plan.

training runs. Zero-shot on TimesFM 2.5.

Integration

Two calls to production

Create a key in the dashboard, then POST your series. That is the whole integration — no SDK to install, no schema to learn beyond one JSON object.

Create a keyIssue a vct_live_ key from the dashboard. Only an HMAC is stored; the plaintext is shown once.

POST your seriesOldest first, evenly spaced. One series or up to 512 of them in a single request.

Read the bandThe median plus nine deciles come back. Plan on the P90, alert on the P10.

Request

bash
curl https://api.vectime.cloud/v1/forecast \
  -H "Authorization: Bearer $VECTIME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "series": [412, 430, 455, 471, 502, 538, 561, 590, 617, 641, 668, 702],
    "horizon": 6,
    "quantiles": true
  }'

Response

json
{
  "model": "timesfm-2.5-200m",
  "horizon": 6,
  "forecasts": [
    {
      "index": 0,
      "point": [728.4, 754.1, 781.9, 806.2, 833.7, 858.0],
      "quantiles": {
        "q10": [701.2, 716.8, 733.4, 748.1, 764.9, 776.3],
        "q50": [728.4, 754.1, 781.9, 806.2, 833.7, 858.0],
        "q90": [757.9, 793.0, 829.6, 866.2, 903.8, 941.1]
      }
    }
  ],
  "usage": { "series": 1, "forecast_points": 6, "requests_this_period": 42 }
}

What you get

Everything around the model, already built

The parts you would otherwise write yourself around a research checkpoint.

Probabilistic outputEvery forecast returns the median plus nine quantiles (P10–P90), so you can size safety stock or set alert thresholds instead of guessing.

Batch in one requestSend up to 512 series per call. They are forecast together in a single model pass, which is far cheaper than a request per SKU.

Metered and rate limitedToken-bucket limits per key and an atomic monthly quota. Both are returned in RateLimit-* and X-Quota-* response headers.

Keys you can rotateIssue and revoke keys per environment. Only an HMAC of each key is stored — the plaintext is shown exactly once.

Usage you can auditPer-request logs with status, latency and horizon, aggregated into daily counts in the dashboard.

Plans that take effect nowLimits are read from your plan on every request, so an upgrade applies to the very next API call.

Benchmarked, not promised

300 M4 series against Prophet and seasonal-naive

We ran the comparison ourselves and published every per-series row — including the third of monthly series where the plain baseline beat us.

9.1sMAPE

TimesFM on M4 Hourly, against 21.2 for Prophet at defaults.

82.5%coverage

Of held-out points inside the 80% band Hourly. Prophet's 80% band covered 65.0%.

34 / 100losses

Monthly series where seasonal-naive beat TimesFM. Run it on your own data first.

Read the full benchmarkPer-series CSV and the script are downloadable(opens in new tab). One command reruns it on a free key.

Plans

Priced by volume. Same weights on every tier.

Calls per month, requests per minute and request size are what change. The model does not.

Free

$0/month

Developer tier — kick the tyres on real TimesFM weights.

Monthly calls300

Rate limit10/min · burst 5

Series / request4

Context window1,024 points

Max horizon64 steps

✓300 forecast calls / month (~10 a day)
✓10 requests / minute, burst 5
✓Up to 4 series per call, horizon 64
✓1,024-point context window
✓Quantile bands (P10–P90)
✓Community support

Start free

Starter

$49/month

For a first production forecasting job.

Monthly calls50,000

Rate limit60/min · burst 30

Series / request32

Context window2,048 points

Max horizon128 steps

✓50,000 forecast calls / month
✓60 requests / minute, burst 30
✓Up to 32 series per call, horizon 128
✓2,048-point context window
✓Quantile bands (P10–P90)
✓Email support, 2 business days

Choose Starter

Pro

Enterprise

Custom

Dedicated capacity, your compliance rules.

Monthly callsUnmetered

Rate limit2000/min · burst 500

Series / request512+

Context window4,096+ points

Max horizon512 steps

✓Unmetered forecast calls
✓2,000 requests / minute, burst 500
✓512+ series per call, batch sizing on request
✓4,096+ point context window
✓Dedicated inference capacity
✓VPC peering & data residency options
✓Audit log export
✓Dedicated support channel, SLA agreed per contract

Talk to sales

Need a dedicated deployment or data residency?Compare plans in detail

Your first forecast is one curl away.

300 forecasts a month on the free tier, quantile bands included, no card. Run the benchmark on your own series before you believe ours.

Start free API reference