TimesFM-3: A zero-shot foundation model for multivariate forecasting

Ayush Jain and Rajat Sen, Google Research, 31 August 2026, with Yichen Zhou, Petros Mol, Abhimanyu Das and Samet Oymak acknowledged. The announcement post for the third generation of timesfm. T3 — a first-party release announcement. The architectural facts in it are the maker describing its own artifact and can be taken at face value; the benchmark claim is the maker grading its own homework, and in this case the post does not even publish the grades (see below).

What changed

Multivariate in the model rather than around it. TimesFM-1 and 2.5 were, in the post's words, "strictly limited to univariate forecasting" — the 2.5 tutorial this wiki read got covariates in through XReg, a fusion step bolted on before or after the forecast (timesfm-2-5-forecasting-tutorial). 3.0 keeps the decoder-only transformer and alternates two attentions: causal temporal attention along time, and full variate attention across series. Cross-series structure is learned in pretraining instead of being patched in at call time. The model takes past covariates, past-and-future covariates (a promotion, a weather forecast) and several targets at once, and returns point and quantile forecasts — nine quantiles, p10 to p90.

Non-autoregressive decoding. Contiguous Patch Masking lets the model emit the entire horizon in a single forward pass rather than generating patch after patch. Series are still cut into 32-step patches and normalised per series. The published card gives 20 transformer layers, model dimension 1280, 16 heads.

Scale of the corpus. "More than 1 trillion time points," real and synthetic.

The benchmark claim, and what it does not say

Three benchmarks: gift-eval, FEV-Bench and a TIME-leaderboard. The claim is that TimesFM-3 is "the top-ranked model in terms of both point and probabilistic forecasting metrics among all pre-trained foundation models," beating Chronos-2, the Toto 2.0 family and TimesFM-2.5, and that univariate mode alone already matches or beats them before covariates are added.

Two things sit under that sentence.

The post ships charts and no numbers. No table, no CRPS, no rank: the comparison is a picture. A wiki that recorded PatchTST's 0.563 and rank 5.72 from gift-eval cannot record anything comparable here, so nothing on this page can be checked against the leaderboard it names.

The comparison class is foundation models only. GIFT-Eval's answer was that the overall winner is PatchTST, a supervised transformer trained per dataset, ahead of every foundation model. "Top among pre-trained foundation models" is compatible with that and does not touch it. So the release does not move this spoke's open question 7, which asks about the per-dataset supervised comparison — and it is worded in a way that avoids the question rather than answering it.

Worth noting where the model aims. GIFT-Eval found the two things pretraining did not buy were noise and cross-series structure — multivariate configurations went to PatchTST and iTransformer, and Moirai was then the only foundation model that would even accept multivariate input. TimesFM-3 is built at exactly that gap. Whether it closes it is measurable, and the post declines to measure it in public.

Availability

• Code: github.com/google-research/timesfm. Weights: google/timesfm-3.0-pytorch on Hugging Face.
• The licence changed. The 2.5 card reads license: apache-2.0; the 3.0 card reads license: other with license_name: timesfm-non-commercial-license-v1.0. Both checked on 2026-09-01. Open weights, commercial use withdrawn, and the announcement post does not mention it.
• BigQuery integration via AI.FORECAST "in the coming weeks", so the release is also a funnel into Google's warehouse.

Related

timesfm · gift-eval · time-series-forecasting · timesfm-2-5-forecasting-tutorial · machine-learning · synthesis