Product case study · September 1, 2026
TimesFM-3 forecasts many series at once — but its default weights are non-commercial
Google's 330M TimesFM-3 handles multivariate forecasting and covariates, but its default PyTorch weights are not licensed for commercial or production use.

Keep downloaded TimesFM-3 weights in permitted non-commercial, non-production evaluations. BigQuery now documents a separate commercial route with multivariate forecasting in Preview. Superbash's fixed three-dataset packet remains unmeasured.
Checked October 29, 2026: official documentation and the existing benchmark packet were reviewed. No model inference or BigQuery query was run for this refresh.
The default TimesFM-3 weights are not licensed for commercial or production use. Google Research’s repository distinguishes its Apache-2.0 code from the timesfm-non-commercial-license-v1.0 on the v3 pretrained weights and says commercial or production use of those default weights is not permitted.[1] That restriction comes before any accuracy discussion: a technically attractive checkpoint is not an approved deployment route.
The exact release is google/timesfm-3.0-pytorch, a 330 million-parameter PyTorch checkpoint for time-series forecasting.[2] Google introduced it on August 31, 2026 as a zero-shot model that can jointly forecast multiple targets and accept past-only plus known-future covariates.[3] The release tag is v3.0.0 at commit 331c6d33cb1ac2611de3056d0ac7164aab6301eb.[4]
Access and license boundary
Hugging Face labels the checkpoint for time-series forecasting and lists the TimesFM Non-Commercial License v1.0.[2] At the October 29 check, its card still said that no Inference Provider deployed this model. That provider listing concerns this downloadable checkpoint; it does not describe every Google-hosted service.[2]
Do not collapse these facts into “Apache-2.0 weights.” The repository code is Apache-2.0, while the v3 pretrained weights are separately restricted.[1] The launch announcement described BigQuery integration as forthcoming.[3] The current BigQuery documentation now lists multivariate forecasting tutorials in Preview and explicitly says TimesFM usage within BigQuery permits commercial use under Google Cloud terms. The downloadable v3 weight license does not restrict that hosted usage.[5]
What Google claims—and what Superbash measured
Google’s benchmark claims
Google says TimesFM-3 ranked first among pretrained foundation models on GIFT-Eval, FEV-Bench, and TIME under its own evaluation setup, and says multivariate mode improves its average rank when cross-series information and covariates are available.[3]
Those are vendor claims, not a Superbash result. They do not establish accuracy on a particular retailer, telemetry system, or planning process.
Superbash evidence
The fixed TimesFM-3 benchmark packet defines three CC BY 4.0 UCI datasets, chronological held-out horizons, a seasonal-naive baseline, MAE, SMAPE, wall time, peak memory, raw-prediction paths, and an at-least-2-of-3 MAE gate. It remains unmeasured. Its September 1 status record describes a worker with no NVIDIA GPU, only 7.5 GiB RAM, and no approved alternate accelerator; those are historical run constraints, not a fresh hardware measurement. No checkpoint was downloaded and no predictions, metrics, plot, or performance result was manufactured.[6]
An unmeasured packet is not a failing score and is not evidence that TimesFM-3 loses to seasonal naive. It is a reproducible protocol plus a documented hardware limit.
The evaluation that remains
The protocol transforms public source data into three distinct forecasting scenarios: retail-like multivariate demand from Online Retail; monitoring-style hourly traffic telemetry from Metro Interstate Traffic Volume; and Bike Sharing, with casual and registered rental counts as targets and calendar/weather columns as declared covariates.[6] Weather observations need the availability check below before being treated as known future inputs. Each source is CC BY 4.0 and remains outside this repository; the packet records attribution, download URLs, transformations, and frozen split rules.[6]
For every dataset, the runner must use the same context and held-out horizon for TimesFM-3 and a seasonal-naive forecast. It must retain raw forecasts and actuals, calculate MAE and SMAPE, record wall time and peak RSS, and reject invalid values. The pass gate is intentionally narrow: TimesFM-3 must beat seasonal naive on MAE in at least two of the three held-out datasets. A mixed or negative outcome remains publishable evidence.
If you have suitable approved hardware, use the pinned v3.0.0 code revision and checkpoint, run only the public data protocol in a non-commercial evaluation setting, and preserve failures. Do not transfer a vendor rank into a production recommendation, and do not convert the Apache-2.0 code license into permission to use the default v3 weights commercially.
Choose an access route before asking an agent to run anything
| Route | What the checked sources establish | Next check |
|---|---|---|
Downloaded google/timesfm-3.0-pytorch | Separate non-commercial, non-production weight license | Read the actual license against the proposed use, including how results will be used. |
| TimesFM within BigQuery | Commercial use under Google Cloud terms; multivariate tutorials marked Preview | Review the selected feature’s Preview terms, permissions, data location, and current charges. |
The weight license defines permitted evaluation more narrowly than “an internal experiment”: its non-commercial definition excludes results used in commercial decisions, client deliverables, or paid services.[7] A self-hosted proof of concept for a client should not be assumed permitted merely because it is small.
For BigQuery, start with Google’s single-series multivariate tutorial. Treat it as a separate service evaluation, not a reproduction of the pinned local benchmark. Record the query and service options, input table snapshot, held-out timestamps, job ID, and observed billing. This article reports neither a tested query nor a price estimate.
Audit future inputs before measuring accuracy
Google distinguishes past-only covariates from inputs available across both history and the forecast horizon.[3] Our evaluation recommendation is to write down when each value became available, rather than assuming every column in a historical CSV was knowable at prediction time.
| Column type | Forecast-time treatment |
|---|---|
| Calendar weekday or published holiday | Can be known ahead of the horizon. |
| Historical traffic or observed weather | Use only observations available by the forecast cutoff. |
| Future weather forecast | Use the forecast issued by the cutoff, retaining its issue time. |
| Weather actually observed during the held-out period | Do not present it as an operationally available future input. |
Future casual or registered rentals | Keep as held-out targets, never input features. |
The existing Bike Sharing packet needs this qualification. Its frozen pastFutureCovariateColumns includes weathersit, temp, atemp, hum, and windspeed. UCI describes the dataset as rental counts with corresponding weather information, not an archive of forecasts issued before each cutoff.[8] A run using those held-out weather observations would be conditional on future weather being known. It would not establish real-world forecasting accuracy with only information available at the time.
Keep that limitation beside any eventual score. For an operational experiment, propose a separately reviewed protocol using only available inputs; do not silently alter the frozen comparison. The benchmark files remain unchanged and unmeasured in this refresh.
A bounded agent task you can review
Use this proposed review prompt with a coding agent that has access to the benchmark checkout. It is an editorial exercise, not a tested agent run:
Read benchmarks/timesfm-3/README.md, protocol.json, and results/status.json.
Do not install packages, download weights/data, run inference, or edit files.
Return a review table with: target, context length, horizon, seasonal period,
prepared-input hash status, and every covariate's availability at the cutoff.
Flag observed future weather and missing evidence. Distinguish the historical
host record from the current machine. Finish with blockers to a measured run;
do not invent an accuracy score or change the unmeasured status.
Verify the response against the files yourself. A useful review must notice the unfilled prepared-input hashes, the missing candidate results, and the Bike Sharing weather assumption. If it misses one, ask for the exact supporting field before allowing execution. Continue with the model hosting decision lesson to record the chosen route and its operating requirements.
Sources
Official sources checked October 29, 2026 unless noted. Benchmark status and protocol were inspected locally on that date; their original run record is dated September 1.
[1] https://github.com/google-research/timesfm/tree/v3.0.0 — TimesFM v3.0.0 repository README and its explicit code-versus-weights license notice
[2] https://huggingface.co/google/timesfm-3.0-pytorch — checkpoint identity, model size, weight license, modality, and checked Inference Provider status
[3] https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/ — Google Research announcement, capabilities, benchmark claims, and forthcoming BigQuery integration
[4] https://github.com/google-research/timesfm/releases/tag/v3.0.0 — v3.0.0 release and commit 331c6d33cb1ac2611de3056d0ac7164aab6301eb
[5] https://docs.cloud.google.com/bigquery/docs/timesfm-model — commercial hosted-use terms and multivariate Preview tutorials; page last updated October 27, 2026
[6] Superbash benchmark packet — local protocol.json, results/status.json, and sources/datasets.json; dataset attribution and license records are pinned there. UCI Online Retail checked October 29; Metro source could not be fetched during this refresh, so its license attribution remains the packet’s September 1 record.
[7] https://huggingface.co/google/timesfm-3.0-pytorch/blob/main/LICENSE — full weight license, including the definition of non-commercial purpose and separate API terms
[8] https://archive.ics.uci.edu/dataset/275/bike+sharing+dataset — rental targets and corresponding weather fields
Put this to work
Separate Apache-2.0 repository code from the separate license on the v3 pretrained weights before evaluating a forecasting stack.
Try
Audit the fixed public-data packet for forecast-time covariate availability before any permitted non-commercial run.
Prove it worked
Keep each chronological split immutable; retain raw predictions, MAE, SMAPE, wall time, peak memory, and failures alongside the seasonal-naive baseline.
Where it can pay
Do not use the default v3 weights in client or production forecasting without a separately permitted route.
Keep in view
- The exact checkpoint is `google/timesfm-3.0-pytorch`: 330M parameters for multivariate time-series forecasting.
- The v3 pretrained weights use TimesFM Non-Commercial License v1.0; Google says commercial and production use of the default weights is not permitted.
- BigQuery documents commercial use under Google Cloud terms; its multivariate forecasting tutorials are marked Preview.