Google's TimesFM-3 forecasts every variable at once
A quiet-but-real model release out of Google Research, and a state-backed number for just how much computation the agent era is expected to eat.
Google's TimesFM-3 became the first model in the family that can forecast many variables at once — and it fills the entire future in a single pass. The 330-million-parameter transformer chops each series into patches of 32 points, normalizes them onto one scale, then alternates two attention directions: along time within a series (past values only, so nothing leaks forward) and across series at each timestamp, which is how it learns that discounting one product moves another. Earlier TimesFM versions predicted one block at a time, slow and error-compounding; TimesFM-3 marks every future step as blank and fills them at once, and Google says that multivariate mode wins the average rank on point and probabilistic forecasting across GIFT-Eval, FEV-Bench and the TIME leaderboard. The write-up is dated August 31 and the weights are open on Hugging Face, so this is a shippable release rather than a teaser — and the fact that a forecasting frontier model fits in 330 million parameters is the part most coverage skipped: for narrow numerical tasks, scale is not the binding constraint. It also lands in a race that is no longer America's — Ant International's Falcon TST 2.0 claimed the same GIFT-Eval crown in late August, in our report Ant's Falcon TST 2.0 tops GIFT-Eval as global banks adopt the model.
China's state telecom research arm now expects the country to consume 100 quadrillion tokens this year, and more than 35 quintillion by 2030. The forecast, published in the China Telecom Research Institute's report on AI infrastructure in the agent era and carried by CCTV News, implies a compound annual growth rate approaching twelve times — a figure its director Rao Shaoyang tied to the industry's shift from model and compute competition to deploying agents at scale and monetizing them. Two secondary numbers matter more than the headline: compute demand is projected to grow roughly tenfold a year for the next two to three years, and inference is expected to reach 80% of China's compute market by 2029. Read together, that is a state-adjacent body telling the market that the token bill, not the training run, is the constraint worth planning around — the same shift we measured from the demand side in AI agents now burn more tokens than humans on OpenRouter. Treat the trajectory as a planning projection, not a measurement: nobody audits token counts across a whole economy.
What to watch: whether anyone outside Google and Ant publishes an independent multivariate forecasting run on TimesFM-3's checkpoints.
If a 330M-parameter model can beat the big labs at forecasting, is scale the right thing to spend on anymore? Tell us in the comments.
