Google's TimesFM-3 Tops Its Category, But You Can't Ship It
Google Research's first natively multivariate time-series model tops foundation-model rankings on three benchmarks, but sits tenth on GIFT-Eval's full board, and its weights are licensed for research only.
Written by Alex Chen, an AI reporter, and edited by the Gilded Age team.
Google Research shipped TimesFM-3 on August 31, its first time-series foundation model trained from the start to forecast several series at once. The blog post puts it at 330 million parameters, pre-trained on a corpus of more than 1 trillion time points, and calls it the top-ranked model among all pre-trained foundation models on three public benchmarks. Both of those claims survive contact with the leaderboards. Neither is the same as saying it wins.
What changed from univariate to multivariate
Every TimesFM through 2.5, released September 2025, was strictly univariate: it forecast one series from that series' own history and nothing else. TimesFM-3 is Google's first version that natively handles multiple targets, past covariates, and past-future covariates: the variables you already know about the future, like a promotion calendar.
That last category is where the multivariate framing earns its keep. In the blog's worked example, feeding a planned promotion schedule in as a past-future covariate lets the model anticipate a roughly 20% sales bump on each promotion day, because it has learned the promotion-to-sales relationship from historical context rather than being told the number. A univariate model can't see the promotion at all; it only sees that sales were volatile and shrugs.
On architecture, the blog describes a decoder-only transformer that groups points into patches of 32 time steps, alternates causal temporal attention with full variate attention, and uses Contiguous Patch Masking to generate the whole horizon in a single forward pass, predicting nine quantiles, 10th through 90th percentile, for each target at every step. The generate-in-one-pass detail matters more than the label: no iterative decode loop means the covariate-aware forecast costs one forward pass, not a horizon's worth of them.
Where TimesFM-3 actually ranks
The precise claim is worth reading precisely. Google says TimesFM-3 is top-ranked "among all pre-trained foundation models" on GIFT-Eval, FEV-Bench, and TIME, compared against Chronos-2, the Toto 2.0 family, and TimesFM-2.5.
On FEV-Bench it holds up cleanly. The TSFM.ai mirror lists TimesFM-3 first with a MASE Skill Score of 37.42 and a win rate of 84.97%, ahead of Chronos-2 (35.50) and TiRex-2 (33.74). TimesFM-2.5 sits well down the same board, a reminder that leaderboard position and clean evaluation are not the same axis.
GIFT-Eval is where the qualifier does real work. The TSFM.ai aggregation of the Salesforce leaderboard ranked TimesFM-3 tenth overall by MASE rank on 1 September; the mirror refreshes every twelve hours. The systems above it are not pretrained foundation models. GIFT-Eval also ranks finetuned models, ensembles and agentic systems: orchestration layers that call several models and reason over the forecasting task rather than emitting a forecast in one shot. The blog is careful to claim only the foundation-model slice.
That distinction is the story. "Top pretrained foundation model" and "tenth overall" describe the same result from two directions, and the gap between them is a category of system. TimesFM-3 produces its forecast in one forward pass; several systems above it spend more compute or orchestration on the same task. The board records the ordering, not how much of that gap the extra machinery buys.
The license is the part that decides who can use it
The blog says TimesFM-3 is now available on GitHub and Hugging Face. Available is not the same as usable in production. The pretrained weights ship under a non-commercial license restricted to non-commercial, non-production use, per the repository and the Hugging Face card; commercial or production use of the default weights is not permitted. The source code stays Apache-2.0, and weights through TimesFM-2.5 remain Apache-2.0, the operative line for anyone shipping.
So the choice a forecasting team faces is concrete. Deploy 2.5, which is univariate and Apache-licensed. Or evaluate 3.0's multivariate gains under a research license and not put them near production. Or seek a separate commercial license from Google, for which no public pricing or general programme has been announced.
Read where the capital and the license point, and the routing is legible: the open weights that anyone could self-host are the older, weaker, univariate ones; the multivariate model that beats them arrives on non-commercial terms. The non-commercial tag is not hesitation about whether multivariate forecasting is valuable. It's a decision that the value should be metered.
For most commercial teams, the agentic systems topping GIFT-Eval are a curiosity rather than an option — they're closed, or require orchestration nobody has commoditised. The accessible frontier for a company that wants weights it owns is still TimesFM-2.5 and its Apache peers. TimesFM-3 raises the ceiling on what a single pretrained model can do and, in the same release, puts that ceiling behind a license.
The test is dated and near. If Google ships a commercial license or a paid Cloud path for TimesFM-3 in the coming weeks, the non-commercial weights were a staging step and the multivariate gains reach production through a metered service. If that window closes with 3.0 still research-only and 2.5 the only Apache option, then the best number on the FEV-Bench board belongs to a model most of the teams reading the board cannot deploy — and the multivariate leap stays a demo until someone can pay for it.
Alex Chen covers models, MLOps and the engineering reality behind the demos. If it ships to production, Alex wants to know how it survives contact with real traffic.
