Google Open-Sources TimesFM 3.0: 330M-Parameter Multivariate Zero-Shot Time Series Foundation Model
On September 1, Google Research released and open-sourced TimesFM 3.0, a next-generation time series foundation model. With 330M parameters and pretraining on over one trillion real and synthetic time points, it is available on Hugging Face (timesfm-3.0-pytorch) and GitHub, with BigQuery integration coming soon.
Confirmed
- The model was released by Google Research; Hugging Face researcher osanseviero helped announce it, with simultaneous availability on GitHub and Hugging Face.
- The biggest improvement is native multivariate forecasting: previous versions (TimesFM-2.5 and earlier) were univariate only, while the new model jointly forecasts multiple co-evolving series zero-shot, capturing cross-series dependencies without task-specific fine-tuning, and generates forecasts in a single forward pass.
- Per @rseroter, the model can also leverage external covariates such as weather and promotions.
- Per @ymatias, TimesFM-3 ranks first on Gift-Eval, FEV-Bench and TIME benchmarks; per @anselm, it substantially outperforms Amazon's Chronos 2 and Toto 2.0 on major benchmarks at lower deployment cost due to its lightweight size.
- BigQuery integration is upcoming (@ymatias).
Why it matters
- Time series forecasting has long relied on task-specific models; TimesFM 3.0's zero-shot foundation-model approach sharply lowers the barrier to adoption, and native multivariate support fixes the key shortcoming of earlier versions.
- The lightweight, open-source release lets smaller teams deploy high-quality forecasting locally, directly challenging Chronos, Toto and other competitors.
2026-09-01 ~ 2026-09-02 · 11 related posts
Primary sources
- Google releases TimesFM-3: Multivariate time series foundation model — SaxenaNayan · 2026-09-01
- Google releases TimesFM 3.0, a 330M open foundation model for time series — osanseviero · 2026-09-01
- Google's 330M TimesFM-3 Beats Rivals in Time Series Forecasting — anselm · 2026-09-01
- Google releases TimesFM-3, a zero-shot foundation model for multivariate forecasting — Recoil42 · 2026-09-01
- [source] Google Releases TimesFM 3.0 for Time-Series Forecasting — google · 2026-09-01
- TimesFM-3: Native Multivariate Zero-Shot Forecasting Model — ymatias · 2026-09-01
5 near-duplicate retellings: osanseviero · aigclink · ymatias · SumitGup · rseroter