Google Ships TimesFM-3, Drops the Apache License It Used Before

Google released TimesFM-3, a time-series forecasting model, on Aug. 31 under a license that blocks commercial or production use, reversing the Apache 2.0 terms of its two predecessors.

The release appeared on Google's own research blog, credited to researchers Ayush Jain and Rajat Sen. Weights are posted on GitHub and Hugging Face under what the model card calls the TimesFM Non-Commercial License v1.0.

TimesFM-3 has 330 million parameters and trained on more than 1 trillion time points of real and synthetic data, according to the blog post. It is the first version pretrained natively for multivariate forecasting, meaning it can predict several related series in one pass instead of one series at a time, which is all versions 1.0 through 2.5 could do.

What changed since 2.5

Google said TimesFM-3 ranked first among pretrained foundation models on three benchmarks: GIFT-Eval, FEV-Bench and the TIME leaderboard. The company also said the model matches or beats Chronos-2 and Toto 2.0 even in single-series mode. Those are Google's own comparisons; neither Chronos-2's maker nor Toto 2.0's maker has published a response.

The license is the sharper break. Hugging Face model cards for both prior releases list a plain "apache-2.0" tag: TimesFM 2.0, at 500 million parameters, and TimesFM 2.5, at 200 million. That puts Google on the opposite path from open-weight releases this month like DeepSeek's V4.1-Flash, which shipped under a fully permissive MIT license.

Version	Parameters	License
TimesFM 2.0	500M	Apache 2.0
TimesFM 2.5	200M	Apache 2.0
TimesFM 3.0	330M	Non-Commercial v1.0

TimesFM-3's own model card states the license restricts use to "non-commercial and non-production" purposes, full stop.

The part Google didn't explain

Google's blog post announcing TimesFM-3 does not mention the license at all, let alone say why a smaller, newer model that beats its predecessors ships under tighter terms than either of them. The restriction appears only on the Hugging Face and GitHub pages, not in the announcement itself.

Google said BigQuery integration for TimesFM-3 is coming "in the coming weeks," while TimesFM-2.5 remains available today through BigQuery's AI.FORECAST function. Until that integration ships, and until Google says whether the non-commercial terms are permanent, anyone forecasting with TimesFM-3 in production is reading the older, Apache-licensed model's documentation and running the wrong one.

The restriction also sets TimesFM-3 apart from smaller open releases this week, including OpenBMB's MiniCPM5-2B, which competes on benchmark scores without gating commercial use behind a separate agreement.

Who actually uses a forecasting model

TimesFM's own pitch is zero-shot forecasting: point it at a new series, sales figures, energy demand, server load, and it predicts without task-specific fine-tuning. That is exactly the kind of internal, revenue-generating use a non-commercial license is written to exclude.

Google's own probabilistic output, nine quantiles running from the 10th to the 90th percentile rather than a single predicted number, is aimed at exactly that kind of production planning, according to the research blog. The company did not say whether it considered a paid commercial tier alongside the free non-commercial one, the way some open-weight labs license a smaller model freely and a larger one for a fee.

For now, a team that wants TimesFM-3's multivariate forecasting in a shipping product has three options: wait for Google to loosen the license, fall back to TimesFM-2.5's weaker but Apache-licensed univariate model, or build the multivariate comparison in-house and confirm whether TimesFM-3's benchmark lead holds on their own data before committing to either path.