TimesFM-3

Why teams choose it

Tradeoffs to know

• Commercial or production use of the default TimesFM-3 pretrained weights is not permitted under the published non-commercial license.
• Benchmark ranks versus Chronos-2, Toto 2.0, and TimesFM-2.5 are vendor-reported on GIFT-Eval, FEV-Bench, and TIME and should be reproduced on your series.

When not to use this

• BigQuery TimesFM-3 integration was announced as coming in the following weeks; do not assume 3.0 is live in AI.FORECAST yet.
• Self-hosting outcomes depend on hardware, quantization, and ops maturity—budget time beyond swapping an API hostname.
• May demand more instrumentation than SaaS-managed APIs to duplicate latency, failover, and support guarantees.

Technical specs

Inputs

time-series

Outputs

time-series

Capabilities

univariate forecasting, multivariate forecasting, zero-shot forecasting, past covariates, past-future covariates, quantile forecasts, single-pass decoding

License

TimesFM Non-Commercial License v1.0 (pretrained weights); Apache 2.0 (code)

Model string

timesfm-3

Benchmarks

{
  "source": "https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/",
  "fev_bench": "top pretrained foundation model",
  "gift_eval": "top pretrained foundation model",
  "parameters": "330M",
  "vendor_reported": true,
  "time_leaderboard": "top pretrained foundation model",
  "pretrain_time_points": "1T+"
}