---
type: analysis
slug: research-timesfm-3-handles-multiple-series-and-covariates-natively-but-its-weights-are
title: "TimesFM-3 handles multiple series and covariates natively, but its weights are now non-commercial and commercial use goes through BigQuery"
description: "Google's 330M-parameter TimesFM-3 forecasts several targets with past and known-future covariates in one forward pass, with no fine-tuning. Unlike TimesFM 2.5, the released weights carry a non-commercial, non-production license, and Google points commercial users to BigQuery ML."
published_at: 2026-10-23T06:00:00.000Z
topics: ["google","research"]
source_urls: ["https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting","https://github.com/google-research/timesfm","https://huggingface.co/google/timesfm-3.0-pytorch","https://docs.cloud.google.com/bigquery/docs/timesfm-model","https://cloud.google.com/blog/products/data-analytics/bigquery-timesfm-3-0-with-multivariate-time-series-forecasting","https://huggingface.co/spaces/Salesforce/GIFT-Eval","https://huggingface.co/spaces/autogluon/fev-bench","https://huggingface.co/amazon/chronos-2"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
On August 31, 2026, Google Research published TimesFM-3, by Ayush Jain and Rajat Sen, as the third major version of its time-series foundation model. Every earlier TimesFM checkpoint was univariate: it forecast one series from that series' own history. Covariates in 2.5 came in through a separate regression add-on (XReg). TimesFM-3 is pretrained to forecast several targets jointly. It also takes past-only covariates and past-and-future covariates, all zero-shot. On October 13, 2026, Google Cloud announced a preview of TimesFM 3.0 inside BigQuery's AI.FORECAST SQL function.

This matters to builders for two reasons. First, real forecasting tasks are usually multivariate: demand alongside promotions, load alongside weather. Zero-shot covariate handling removes much of the feature engineering and per-dataset training these tasks needed. Second, the release changes how Google distributes the model. The weights are no longer freely usable in production, and that affects anyone who built pipelines on Apache-2.0 TimesFM 2.5.

## What we checked

- **Does the TimesFM-3 cluster point to a broader market or architecture shift that one release note can't capture?** partly confirmed. There are two shifts here. On architecture, time-series foundation models are moving from univariate to natively multivariate, with covariates handled inside the model. Amazon's Chronos-2 did this first in October 2025, so TimesFM-3 is Google catching up and claiming the top rank, not starting the trend. On distribution, Google moved TimesFM from Apache-2.0 weights to non-commercial weights with commercial use routed through BigQuery. That is a licensing change, and a single release note doesn't show it. [1][2][8]
- **(Added) Can teams self-host the TimesFM-3 weights in commercial production, as they could with TimesFM 2.5?** refuted. No. The README says the TimesFM 3.0 weights are under the TimesFM Non-Commercial License v1.0, 'restricted to non-commercial, non-production use.' Weights up to 2.5 and all source code stay Apache-2.0. Commercial use is offered through Google Cloud services such as BigQuery ML. [2][3][4]
- **(Added) Do the 'best on GIFT-Eval, fev-bench and TIME' claims hold up independently?** unresolved. Google reports average-rank plots without numeric scores. The GIFT-Eval and fev-bench leaderboard Spaces did not load leaderboard data during this session, so I could not confirm the ranking. [1][6][7]

## What we found

- Specs: 330M parameters, up from 200M in TimesFM 2.5. It is a decoder-only transformer trained on more than 1 trillion time points, real and synthetic. Inputs are cut into 32-step patches. Layers alternate between causal attention over time within each series and full attention across variables. Contiguous Patch Masking lets it produce the whole horizon in one forward pass, with 9 quantiles (10th to 90th percentile) at each step. [1][2]
- The Hugging Face card lists 20 layers, model dimension 1280, 16 heads and an output patch length of 64. Training data includes GiftEvalPretrain (excluding datasets that overlap with fev-bench), Wikipedia pageviews up to November 2023, Google Trends up to the end of 2022, and synthetic data. The README says contexts longer than 15,360 points are truncated to the most recent points. [3][2]
- License: the TimesFM 3.0 weights are under a separate 'TimesFM Non-Commercial License v1.0', restricted to non-commercial, non-production use. The code and all weights up to 2.5 remain Apache-2.0. The README directs commercial and production use to authorized Google Cloud services such as BigQuery ML. The BigQuery docs say use inside BigQuery falls under the Google Cloud Terms of Service and allows commercial use. [2][3][4]
- In BigQuery, multivariate forecasting is in preview. You pass MODEL => 'TimesFM 3.0' to AI.FORECAST along with TARGET_COLS, PAST_COVARIATE_COLS and FUTURE_COVARIATE_COLS. Future rows hold covariate values with NULL targets. Google says AI.DETECT_ANOMALIES and AI.EVALUATE accept TimesFM 3.0 only for univariate series. Neither the post nor the model doc gives pricing or quotas. [5][4]
- In the open-source API, predict_batch takes targets shaped (num_variates, context_len), past-only covariates shaped (k, context_len), and past-future covariates shaped (k, context_len + horizon). A 3-variate example with horizon 24 returns forecasts of shape (3, 24) and quantiles of shape (3, 24, 9). [2]
- Google says TimesFM-3 has the best average rank among pretrained foundation models on GIFT-Eval, fev-bench and TIME, for both point and probabilistic metrics, in multivariate mode. It says univariate mode matches or beats Chronos-2, Toto 2.0 and TimesFM-2.5. The blog shows rank plots, not numeric scores, and I could not load either public leaderboard to confirm. [1][6][7]
- The main open competitor is Amazon's Chronos-2, from October 2025. It is a 120M encoder-only model with Apache-2.0 weights and native support for multivariate data and past and known-future covariates, real and categorical. Its limits are 8,192 context steps and a 1,024-step horizon. So permissively licensed multivariate zero-shot forecasting existed about ten months before TimesFM-3. [8]

## What it means

For prototyping and research, TimesFM-3 is worth testing on any multi-series problem where covariates such as promotions, prices or weather drive the outcome. One forward pass returns jointly forecast targets with quantiles, so you can skip the XReg add-on that 2.5 used for covariates. For production, check the license before anything else. You cannot legally self-host the 3.0 weights in a commercial pipeline. Your options are: (a) use BigQuery AI.FORECAST, accepting a preview feature, Google Cloud lock-in and pricing Google hasn't spelled out for this mode; (b) stay on TimesFM 2.5 (Apache-2.0, univariate plus XReg); or (c) use Apache-2.0 Chronos-2, which already covers multivariate and covariate forecasting.

Run your own backtests before switching. Google's ranking claims show up only as rank plots, and the public leaderboards could not be checked here. The bigger signal for platform planning is the licensing change. A leading lab now keeps its best forecasting checkpoint for its own cloud and leaves the older generation open. If more labs follow, open time-series models could drift behind hosted ones. Build your pipeline so you can switch forecasting backends without rewriting it.

## Still open

- Do the independent GIFT-Eval, fev-bench and TIME leaderboards confirm TimesFM-3's top rank, and by how much?
- What are BigQuery pricing, quotas and the general-availability date for TimesFM 3.0 multivariate forecasting? The Cloud blog says preview, while the README says the BigQuery rollout finished in September 2026.
- What exactly does the TimesFM Non-Commercial License v1.0 allow, for example internal evaluation, academic publication or derivative fine-tunes?
- Will Google offer a commercial license for self-hosting or a Vertex AI endpoint outside BigQuery?

## Sources

1. [TimesFM-3: A zero-shot foundation model for multivariate forecasting (Google Research)](https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting) (primary)
2. [google-research/timesfm GitHub README](https://github.com/google-research/timesfm) (primary)
3. [google/timesfm-3.0-pytorch model card](https://huggingface.co/google/timesfm-3.0-pytorch) (primary)
4. [BigQuery TimesFM model documentation](https://docs.cloud.google.com/bigquery/docs/timesfm-model) (primary)
5. [BigQuery TimesFM 3.0 with multivariate time-series forecasting (Google Cloud blog)](https://cloud.google.com/blog/products/data-analytics/bigquery-timesfm-3-0-with-multivariate-time-series-forecasting) (primary)
6. [GIFT-Eval leaderboard (Hugging Face Space)](https://huggingface.co/spaces/Salesforce/GIFT-Eval)
7. [fev-bench leaderboard (Hugging Face Space)](https://huggingface.co/spaces/autogluon/fev-bench)
8. [amazon/chronos-2 model card](https://huggingface.co/amazon/chronos-2)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- topic hub: [Research](https://newruntime.com/topics/research.md): Explore the research topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)