Google TimesFM-3: 330M Parameter Multivariate Forecasting Model, Non-Commercial Only

Google Research released TimesFM-3 on August 31, 2026, a 330 million parameter time series forecasting model pretrained on more than 1 trillion time points. It is the first TimesFM version built natively for multivariate forecasting, meaning it can process multiple related data series, along with historical and future context variables, in a single pass with no task-specific fine-tuning. The model tops the leaderboards on GIFT-Eval, fev-bench, and TIME, but the weights are restricted to non-commercial, non-production use only.

What happened

Detail	Value
Model name	TimesFM-3
Parameters	330 million
Pretraining data	1 trillion+ time points (real and synthetic)
Patch size	32 steps per patch
Quantile outputs	9 quantiles (10th to 90th percentile) per horizon step
Code license	Apache-2.0
Weights license	timesfm-non-commercial-license-v1.0 (no production use)
Benchmark rank	#1 on GIFT-Eval, fev-bench (100 tasks), TIME (50 datasets, 98 tasks)

Every TimesFM checkpoint up through version 2.5 was univariate: you fed it one series, built from its own history, with nothing else in the picture. That works for simple cases but most real forecasting problems are not shaped that way. Google’s own example is ice cream sales, where foot traffic, weather, promotions, and related product sales all move the target at the same time.

TimesFM-3 handles three input types zero-shot, with no additional training required:

• Multiple targets forecast jointly, each receiving point and quantile outputs.
• Past covariates, variables whose future values are unknown, such as historical foot traffic.
• Past-future covariates, variables whose future values are already known, such as a promotion calendar or holiday schedule.

How the architecture works

The backbone is a decoder-only transformer, the same general family behind large language models. Contiguous data points are grouped into patches of 32 steps and normalized per series, so that a series measured in millions does not drown out one measured in single digits.

Two alternating attention mechanisms handle the forecasting problem:

• Causal temporal attention runs horizontally across time within a single series. It is strictly causal, meaning it cannot peek at future data from the same series.
• Full variate attention runs vertically across all series at a given time step, letting the model learn which series move together.

For past-future covariates, a lookahead trick concatenates the current patch with future patches before tokenization, so known upcoming events (a sale, a bank holiday) are visible to the model before they happen.

One pass for the whole horizon

Earlier TimesFM versions decoded the forecast one patch at a time, which adds latency and compounds errors across longer horizons. TimesFM-3 uses Contiguous Patch Masking, a training strategy first introduced with TiRex. Masked placeholder tokens are appended for the entire forecast horizon upfront. The attention layers fill all of them simultaneously, producing 9 quantiles at every step in a single forward pass.

Why it matters

Multivariate zero-shot forecasting is genuinely useful for businesses with interconnected data: retailers tracking dozens of SKUs, advertisers modeling spend and conversion together, or operations teams forecasting demand alongside supply signals. Until now, applying a foundation model to those problems meant either training a custom model or running separate univariate forecasts and stitching results together.

TimesFM-3 offers a faster path, at least for experimentation. According to Google, it ranks first among pretrained foundation models on all three major benchmarks it was evaluated on, beating Chronos-2, the Toto 2.0 family, and its own predecessor TimesFM-2.5, on both point accuracy and probabilistic calibration metrics.

For teams building production forecasting pipelines, the license wall is the key constraint right now. The repository code is Apache-2.0, but the model weights that actually make it work ship under a non-commercial, non-production license. You can benchmark it and study its outputs, but you cannot serve it behind an API that touches a real business workflow. TimesFM 2.5 remains the only version available for commercial deployment.

Teams exploring AI integration for demand planning or operational forecasting should keep an eye on whether Google releases a commercial-licensed version of the weights. Until then, this is a research and evaluation tool, not a shipping-ready solution. If you are weighing options, our AI integration services cover model selection and deployment alongside the infrastructure decisions that go with it.

Our take

The architecture choices here are solid. Alternating causal and variate attention is a clean way to handle cross-series dependencies without collapsing the temporal structure that makes forecasting hard. The single-pass horizon decoding is a real improvement over autoregressive patch decoding, not just a marketing claim.

That said, benchmark ranks on GIFT-Eval and fev-bench tell you how a model performs on standardized test sets, not on your specific data distribution. Real-world forecasting accuracy depends heavily on feature engineering, data quality, and how well the training distribution matches your domain. We would treat the leaderboard numbers as a signal to run your own evaluation, not as a guarantee of production performance.

The non-commercial license is a meaningful blocker for most businesses. Google has a history of eventually opening weights further, but there is no timeline stated. If you need a deployable multivariate foundation model today, TimesFM 2.5 under Apache-2.0 is still the only option from this family. For a broader picture of what is happening in applied AI tooling, our AI news coverage tracks model releases and licensing developments week by week.

What to do about it

1. Pull the TimesFM GitHub repo and run it against a sample of your own data to get a realistic read on accuracy before any commercial decision.
2. Check whether your use case fits the non-commercial license, internal research and academic work are fine; anything customer-facing or revenue-generating is not.
3. If you need a production-ready option today, evaluate TimesFM 2.5 (Apache-2.0) or Chronos-2 as the nearest alternatives.
4. Watch for a commercial license announcement. When it comes, the single-pass multivariate architecture makes this a strong candidate for retail, advertising, and operations forecasting workflows.

Bottom line: TimesFM-3 is worth benchmarking now if forecasting is a real part of your stack. Just do not plan a production launch around it until the license situation changes.

Source: Marktechpost