Google AI releases TimesFM-3, a 330M parameter model for multivariate time series forecasting

Google Research has released TimesFM-3, a 330 million parameter zero-shot foundation model designed for multivariate time series forecasting. Pretrained on over 1 trillion time points, TimesFM-3 uses a decoder-only transformer with alternating causal temporal and full variate attention to model cross-series dependencies. It ranks first among foundation models on GIFT-Eval, fev-bench, and TIME. While the repository code is Apache-2.0, the TimesFM 3.0 weights are restricted to non-commercial use.

TimesFM-3 multivariate forecasting model

• ▪Google Research released TimesFM-3, a 330 million parameter time series foundation model pretrained natively for multivariate forecasting on more than 1 trillion time points.
• ▪TimesFM-3 forecasts multiple related series in a single forward pass, whereas all previous TimesFM checkpoints through version 2.5 were strictly univariate.
• ▪TimesFM-3 supports three zero-shot input types without task-specific fine-tuning: multiple targets, past covariates, and past-future covariates.

Patch-based transformer architecture

• ▪Through Contiguous Patch Masking, TimesFM-3 generates point outputs and 9 quantiles, from the 10th to the 90th percentile, at every horizon step.
• ▪The decoder-only transformer backbone of TimesFM-3 groups contiguous points into patches of 32 steps and normalizes them per series to prevent scale dominance.
• ▪TimesFM-3 utilizes Contiguous Patch Masking to append masked placeholder tokens for the entire horizon, generating the complete forecast horizon in a single forward pass.

Dual attention mechanisms

• ▪Tokens in TimesFM-3 enter a 2D grid and pass through alternating causal temporal attention and full variate attention mechanisms.
• ▪The causal temporal attention in TimesFM-3 runs horizontally and is confined to earlier tokens inside the same series to block data leakage.
• ▪The full variate attention in TimesFM-3 runs vertically, enabling a token at a given time step to read every other series at that step.

Benchmark leaderboard rankings

• ▪TimesFM-3 achieved the top average rank among pretrained foundation models on the GIFT-Eval, fev-bench, and TIME leaderboards for point and probabilistic metrics.
• ▪TimesFM-3 ranked first overall on fev-bench across 100 real-world tasks and first overall on the TIME leaderboard across 50 domain datasets and 98 evaluation tasks.

Non-commercial licensing restrictions

• ▪TimesFM 2.5 remains the Apache-2.0 licensed option for production deployment, as TimesFM 3.0 weights cannot be deployed behind a production forecast API.
• ▪While the TimesFM repository code is licensed under Apache-2.0, the TimesFM 3.0 weights are restricted to non-commercial, non-production use.

1 source