TimesFM 2.0

TimesFM 2.0 is the second open checkpoint of Google Research's foundation model for time-series forecasting, launched in the project repository on 30 December 2024. Like the rest of the line it forecasts zero-shot: it is shown a history of measurements and returns a horizon of any length, with no training on the user's own data. Two things changed against version 1.0, and Google states both plainly: the checkpoint is up to 25 per cent better on leading benchmarks, and the maximum context is four times longer — 2,048 points of history instead of 512. The gain was bought with size. Where 1.0 had 200 million parameters in 20 layers, 2.0 has 498.8 million in 50 layers, with the model dimension unchanged at 1,280 and history still read in patches of 32 points at a base horizon of 128. The output is still a point forecast: the model ships ten quantile heads, but Google notes they were not calibrated after pretraining, so probabilistic forecasting is marked experimental. The frequency indicator remains mandatory — before forecasting, the series has to be declared as high, medium or low frequency. The training set is the 1.0 corpus plus twenty additional public collections taken from LOTSA, among them Azure virtual machine traces, Borg cluster telemetry, London smart meter readings, ERA5 and CMIP6 climate reanalyses and air quality series from Beijing and across China. This is the checkpoint on which the line first won an independent comparison: after being added to GIFT-Eval, one of the broadest time-series benchmarks available, TimesFM 2.0 took first place on both aggregated MASE and aggregated CRPS, ahead of the next best model by six per cent on MASE. It also remains the workhorse of the family — 86,266 downloads of the PyTorch checkpoint in the thirty days to 17 September 2026, against 22 for the JAX version published ten days earlier. Its weights are Apache 2.0, like 1.0 and 2.5, which matters because the next generation but one, TimesFM 3.0, moved to a non-commercial licence.

▸Videos

No videos yet.

▸Reviews