TimesFM: A Decoder-Only Foundation Model for Time-Series Forecasting
Google Research pretrained decoder-only time-series foundation model (TimesFM) for zero-shot forecasting; ICML 2024 paper. Version 3.0 adds multivariate and covariate support. (Paper with Apache-2.0 code; v1.0–2.5 weights are Apache-2.0, v3.0 weights are non-commercial, non-production)
Verdict
Useful zero-shot baseline for multi-domain series before training a bespoke forecaster. TimesFM 3.0 claims rank #1 on fev-bench, TIME Benchmark, and GIFT-Eval. Available in BigQuery ML and Vertex AI Model Garden for commercial use. Open version not officially supported by Google; v3.0 weights require non-commercial license (production use needs Google Cloud).
License
Custom. Code: Apache-2.0. Weights: v1.0–2.5 Apache-2.0; v3.0 non-commercial, non-production only (timesfm-non-commercial-license-v1.0); commercial use of 3.0 via Google Cloud (e.g. BigQuery ML). Paper: CC BY 4.0.