By GoogleLearning Objectives
- Understand what a time-series foundation model does that classical forecasting does not
- Understand why the 3.0 weights cannot be used commercially while the 2.5 weights can
- Evaluate zero-shot forecasting against the baselines it actually replaces
What Is TimesFM 3.0?
TimesFM 3.0 is a time-series foundation model from Google Research. It forecasts the future values of a numeric series without being trained on that particular series first, the way a language model writes about a topic it was never fine-tuned on. It carries 330 million parameters and was pre-trained on a corpus of real-world and synthetic data comprising more than one trillion time points.
The headline change over TimesFM 2.5 is native multivariate forecasting. Version 2.5 was strictly univariate, taking one series in and returning one forecast out. Version 3.0 forecasts several related series at once and accepts covariates, meaning other variables that help explain the thing being predicted. It handles past-only covariates and also past-future covariates, which are values already known for the forecast window such as a scheduled promotion or a published weather forecast. It decodes non-autoregressively using a technique Google calls Contiguous Patch Masking, emitting the whole forecast horizon in a single forward pass rather than stepping through it one interval at a time.
⚠️Warning
The 3.0 weights are not open source, and this is the most important thing to know before building on them. Google's own repository states that model weights up to version 2.5 remain Apache 2.0, while the TimesFM 3.0 pretrained weights are distributed under a separate TimesFM Non-Commercial License version 1.0, and that commercial or production use of the default pretrained weights is not permitted. The source code stays Apache 2.0 in both cases. So if you need a commercially usable time-series foundation model from this family today, the answer is version 2.5, not version 3.0.
💡Key Concept
Zero-shot forecasting: a foundation model is pre-trained across a very large number of unrelated series, so it can forecast a series it has never seen. That removes the per-series model fitting that classical forecasting requires, which is the real operational saving.
📝Note
Benchmarks here are vendor-reported. Google reports TimesFM 3.0 as the top-ranked foundation model on GIFT-Eval for both point and probabilistic metrics, and as achieving the best average rank on FEV-Bench and on TIME, outperforming Chronos-2 and Toto 2.0. These are Google's own published results and have not been independently reproduced here.
🎯Tip
Find TimesFM: the model card lives on Hugging Face and the code on GitHub. Read the weights license before you plan a deployment.
Pricing
There is no price and no paid tier. The cost is a licensing constraint rather than a bill, and the constraint differs by checkpoint.
- Research and evaluation only
- Non-commercial license version 1.0
- Commercial and production use not permitted
- Apache 2.0
- Commercial use permitted
- The fallback when you need production rights
- Apache 2.0
- Commercial use permitted
- Covers the code, never the weights
Core Features
Native Multivariate Forecasting
Forecasts several related series together rather than one at a time, which is the structural break from version 2.5.
Covariate Support
Accepts past-only covariates and past-future covariates, so known upcoming events can inform the forecast.
Single-Pass Decoding
Contiguous Patch Masking emits the full horizon in one forward pass instead of generating it step by step.
Point and Quantile Output
Returns probabilistic forecasts alongside point estimates, which is what capacity and inventory planning actually needs.
Strengths
- Zero-shot across series — no per-series model fitting
- Multivariate by design — related series and covariates in one model
- Compact — 330 million parameters is small enough to self-host
- Probabilistic output — quantiles, not just a single predicted line
- Open code — the repository itself is Apache 2.0
Limitations and Considerations
- The 3.0 weights bar commercial use — production deployments need version 2.5
- Benchmarks are vendor-reported — Google's own published results
- BigQuery integration is not live — Google describes it as coming
- Forecasting is not causation — covariates improve fit, not explanation
- Compare against the right baseline — the rival is ARIMA and gradient-boosted trees, not a language model
Best Use Cases
| Use Case | Why TimesFM Matters | Caveat |
|---|---|---|
| Many series, no time to fit each | Zero-shot removes per-series modeling | Weights license blocks commercial use |
| Demand planning with known events | Past-future covariates are native | Version 2.5 lacks multivariate support |
| Evaluating foundation models for forecasting | Published results on three benchmarks | Figures are vendor-reported |
| Research and teaching | Free weights and open code | Non-commercial license on 3.0 |
Key Takeaways
- TimesFM 3.0 is Google Research's time-series foundation model, carrying 330 million parameters and pre-trained on more than one trillion time points
- The advance over version 2.5 is native multivariate forecasting with past and past-future covariates, decoded in a single forward pass
- The source code is Apache 2.0, but the 3.0 weights carry a non-commercial license that forbids commercial and production use
- Anyone who needs production rights today should use the version 2.5 weights, which remain Apache 2.0
- Its benchmark results are Google's own, and the fair comparison is against classical forecasting baselines rather than against language models
