Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-09-21. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (open source | 2026-09-21 |
| 2 | 3.0 weights restricted to non-commercial use) | 2026-09-21 |
One-liner: TimesFM (Time Series Foundation Model) is a pretrained model from Google Research that forecasts future values of a numeric series — sales, demand, traffic, sensors — without being trained on your data first. You hand it historical points, it returns predictions.
Verified facts (checked 2026-09-21)
| Attribute | Value | Source |
|---|---|---|
| GitHub stars | 33,362 | google-research/timesfm via GitHub API |
| Repository code license | Apache-2.0 | GitHub API / repo LICENSE |
| Pretrained weights, v3.0 | TimesFM Non-Commercial License v1.0 — non-commercial, non-production use only | LICENSE file in google/timesfm-3.0-pytorch on Hugging Face |
| Pretrained weights, up to v2.5 | Apache-2.0 | Repo README license notice |
| v3.0.0 release | Published 2026-08-28 on GitHub; press coverage followed on 2026-08-31 | GitHub Releases API; MarkTechPost/explainx coverage |
| Parameters | 330 million | HF model card; repo README |
| Pretraining scale | 1 trillion+ time points | Vendor/secondary reporting (Google announcement coverage) |
| Architecture | 20-layer decoder-style transformer, model dim 1280, 16 heads | HF model card |
| Context length | 16,384 | HF model card / README |
| Outputs | Point forecast plus 9 quantiles (10th–90th percentile) | Secondary technical coverage |
| PyPI package | timesfm (PyTorch or Apple-silicon MLX extras: timesfm[torch], timesfm[mlx]) |
Repo README |
| Last repo push | 2026-09-15 (active MLX parity and cache fixes through September) | GitHub API |
The license split is the real story
Until version 2.5, TimesFM had a simple story: Apache-2.0 code, Apache-2.0 weights. Anyone could run, modify, and ship it commercially.
TimesFM 3.0 splits that in two:
- The source code stays Apache-2.0. The repo’s own license notice says so explicitly.
- The 3.0 pretrained weights are released under the
timesfm-non-commercial-license-v1.0. The license text grants rights for “non-commercial and non-production use” only. Commercial or production use of the default 3.0 checkpoint is, in the repo README’s own words, not permitted. - Older checkpoints are unaffected. Weights up to 2.5 remain Apache-2.0, and the repo README states this directly. Pinning to the 2.5 checkpoint keeps the old permissions.
The wording “for the time being” in the repo notice suggests Google may revisit the terms, but nothing more permissive exists as of this check.
What 3.0 actually adds
According to the repo README, 3.0 is the first TimesFM built for native multivariate forecasting — predicting several related series jointly in one pass (2.5 and earlier were univariate). It also accepts covariates natively: past-only signals like historical foot traffic, and past-and-future signals like a promotion calendar whose future dates are known.
The README reports rank #1 overall on three forecasting benchmarks — fev-bench (100 tasks), TIME (98 tasks across 50 datasets), and GIFT-Eval — for foundation models. These are vendor-reported results from the project’s own README; we have not independently reproduced them.
What the marketing doesn’t say
- “Open source” now needs a footnote. Describing TimesFM 3.0 flatly as “Apache-2.0 and free for commercial use” is wrong for the weights. The code is open; the newest checkpoint is not commercially usable as a download.
- Benchmark ranks are self-reported. The #1 claims come from Google’s README and announcement coverage. Benchmark choice and leaderboard framing matter in forecasting, so treat them as strong evidence, not neutral certification.
- The commercial route runs through Google’s hosted products. TimesFM is available in BigQuery ML (SQL), Vertex AI Model Garden (containerized endpoints), and Google Sheets via BigQuery-connected forecasting. Usage there is governed by the relevant Google Cloud terms — and metered like any other cloud service, not free like the download.
- Self-hosting 2.5 commercially means staying on the older model. That is a legitimate, fully Apache-2.0 path, but you give up multivariate support and the 3.0 accuracy claims.
Who should use it
- Data and operations teams prototyping demand, inventory, traffic, or sensor forecasts with no training budget: download the 3.0 checkpoint for internal research freely.
- Teams that want an Apache-2.0 model they can actually ship: pin 2.5 weights or evaluate whether the hosted BigQuery/Vertex route fits.
- Engineers building forecasting into a commercial product on the downloaded 3.0 weights: this is the group most exposed — the license forbids exactly that.
Repository stats, release history, and license text verified against primary sources (GitHub API, the project README, and the Hugging Face model card/LICENSE) on 2026-09-21. This page follows an official-sources-only review: no first-hand benchmark reproduction or deployment was performed, so no hands-on claims are made.