TimesFM wins three benchmarks. Its license may not let you use the answer.

Google published a forecasting model that tops three benchmarks, and licensed it so you cannot use its forecasts in a client deliverable or a production system without a separate commercial license. Ten days after the weights appeared on Hugging Face, its own BigQuery documentation was stamped with an update that says the version you can actually call tops out one release earlier and is comparable to ARIMA. Each page links to the other, and neither mentions the other's version or license terms.

#What TimesFM is

TimesFM is Google Research's time-series forecasting model: point it at a sequence of numbers (revenue, demand, latency, anything with a timestamp) and it predicts what comes next, zero-shot, with no retraining on your data. The current release, TimesFM-3, is a 330-million-parameter transformer; the repository's latest commit as we read it is 8cb0628, dated September 9, 2026, and the 3.0 checkpoint appeared on Hugging Face on August 24. It is not a chatbot wrapper around a forecast; it is the forecasting model itself, and Google is not shy about the results.

#The three benchmark claims hold

We went looking for the overclaim in TimesFM-3's README and did not find one. All three medals are real, re-derived from the live boards rather than taken on the README's word:

• fev-bench: TimesFM-3 is the top row, 87.2% win rate and a 48.7 skill score, ahead of Chronos-2 (82.1/47.3) and TiRex-2 (78.9/45.5).
• TIME Benchmark: TimesFM-3 is the top row on the live leaderboard, 0.64 normalized MASE and 0.536 normalized CRPS, ahead of Toto-2.0-2.5B (0.642), Chronos-2 (0.662), and its own predecessor TimesFM-2.5 (0.669).
• GIFT-Eval: the README calls this one "rank #1 among all foundation models," and that phrase is doing real work. On the full 130-model board, TimesFM-3 sits 13th overall by MASE rank. The twelve entries above it are all labeled "Agentic" systems, not foundation models on their own. TimesFM-3 is the first non-agentic entry, which makes the claim accurate and also a different claim from the other two.

That distinction is the whole lesson in miniature: a reader skimming three gold medals in a row will not notice that the third one changed its own denominator. It is not a lie. It is precise hedging that reads like a uniform boast, and the only way to catch it is to open the leaderboard's detail page instead of stopping at the headline.

#What the license actually restricts

TimesFM-3's weights ship under the TimesFM Non-Commercial License v1.0, and it goes further than the "non-commercial" label suggests. Most licenses in this category restrict redistributing the weights. This one restricts what you do with the answer:

"Non-Commercial Purpose" means use for testing, evaluation, or research not tied to commercial gain, production deployment, or revenue generation. This includes internal benchmarking, academic research, and experimentation on private or public datasets, provided the results are not used in commercial decision-making, client deliverables, or paid products/services.

For clarity, use (a) for any revenue-generating activity, (b) in direct or indirect interactions with end users or production systems, or (c) to train, fine-tune, or distill other models for commercial use, in each case is not a Non-Commercial Purpose.

Read that against how forecasting actually gets used. A consultancy that runs TimesFM-3 on a client's demand data and puts the forecast in a deliverable is outside the license, even if nothing about the weights themselves gets redistributed. The license does say a commercial grant can be requested, at Google's sole discretion; nothing in the repo says what one costs or whether any have been issued. The restriction travels with the output, not just the model file.

Two more details a buyer will miss on a skim: the grant is explicitly "revocable," and distribution of the weights is barred outright, not conditioned. This is also new. The repository's code is Apache-2.0, and every earlier weight release through TimesFM-2.5 stayed Apache-2.0. Only the 3.0 checkpoint carries the new terms, and it has still been downloaded roughly 800,000 times on Hugging Face.

#The surfaces you can actually call say something different

Here is the gap. If you go looking for TimesFM inside Google's own products instead of its GitHub repo, you land somewhere quieter.

BigQuery ML's built-in TimesFM model is described, as of Google's own documentation update on September 3, 2026, as univariate, and the page says plainly that its results are "comparable to conventional statistical methods such as ARIMA." If you want more tuning than it offers, the same page points you to ARIMA_PLUS instead. That overview page never names a version. The AI.FORECAST reference one click away, carrying the same September 3 stamp, does: "Supported models include TimesFM 2.0 and TimesFM 2.5. The default value is TimesFM 2.5," with a note recommending 2.5 for all new forecasting tasks. Google Sheets' forecasting feature, announced in February, runs through BigQuery ML and inherits that ceiling.

So this is not a matter of the product side being vague. It is affirmative: the surface you can call in production tops out at 2.5, and the model that holds the three medals is not on it. Apache-2.0 weights through 2.5 mean you can also run that same generation yourself, outside BigQuery, fully commercially licensed. What you cannot get through BigQuery or Sheets is TimesFM-3. (Google's third surface, Vertex Model Garden, sits behind a login; the public deployment notebooks for it cover 1.0, 2.0, and 2.5, and we did not read the card itself.) None of this is concealment. The benchmark claim and the version disclosure simply live on different pages, and only one of them is the page anybody quotes.

#The agent path runs an older model and never mentions any of this

This is the sharpest version of the gap, and it is the one that matters if you are wiring TimesFM into an agent rather than a notebook. The repository's AGENTS.md tells you to copy its skill directly into your agent's tools: cp -r timesfm-forecasting/ ~/.claude/skills/. Open the skill that instruction installs and it targets TimesFM-2.5, not 3.0: a dozen lines reference 2.5, its version table marks 2.5 as "Latest," and TimesFM-3 appears nowhere in the file (the one "3.0" in it is a z-score threshold). That is at a commit whose README headline is 3.0 and whose own "Archived Model Versions" list contains 2.5. The word "non-commercial" appears zero times. That silence is technically correct, because 2.5 is Apache-2.0 and carries no such restriction. It is also exactly the gap that would mislead someone who later swaps in the 3.0 checkpoint expecting the same freedom.

Stack the three facts: the README you read cites 3.0's benchmark wins, the skill you install runs 2.5's behavior, and the skill itself is community-authored executable Python that a first-party Google file instructs you to copy straight into your agent. We have written before about what it means to install an agent skill without treating it as untrusted code; a first-party research repo telling you to cp -r someone else's Python into your tools directory is that exact pattern, not a hypothetical one.

#The leaderboard may not mean what you think

We went into this expecting to write a section on when a tuned seasonal baseline still beats a foundation model on business data. The evidence did not cooperate, and we would rather say so than write it anyway. The strongest practitioner comparison we found runs the other way: a consultancy that sells AI work reports that in every dataset it tested a time-series foundation model beat Auto ARIMA, citing MAE reductions above 15% on monthly parts sales and about 25% on daily restaurant visitors. Read the table under that sentence, though, and the monthly win came from a fine-tuned model, not a zero-shot one: on the parts data, zero-shot TimesFM scored 0.522 MAE against Auto ARIMA's 0.53, a 1.5% edge, while the 27% win on restaurant visitors was TimesFM's. It publishes no data or code, so treat it as a vendor claim pointing in the vendor's direction. But the TimesFM in that table is v2.0, one of the two generations BigQuery serves, which makes the parts row a second, independent reading of the version an operator actually gets: Google's own doc says comparable to ARIMA, and a practitioner's test of that generation says 1.5% better. That is the closest thing to baseline parity we found, and it came from the vendor's own table rather than from a critic.

What is supported is a critique of the instrument rather than the model. An arXiv paper on TSFM evaluation (Meyer, Kaltenpoth, Zalipski, and Müller, arXiv 2510.13654, third version February 2026) identifies two kinds of information leakage in existing TSFM benchmarks, train-test sample overlap from reusing the same public datasets and temporal overlap between correlated train and test series, and says that ignoring them "risks producing overly optimistic performance estimates that fail to generalize to real-world settings." A benchmark score is a measurement of a model against a test set; if the test set was in the training corpus, the score measures memory.

There is a peer-reviewed result on TimesFM and simple baselines, and its scope matters as much as its finding. An ICLR 2026 poster (Zhu, Carpentier, and Verbeke, "When Foundation Models are One-Liners") examined five families of time-series foundation models, TimesFM among them, and found that for anomaly detection, "performance does not significantly differ to simple one-liner baselines: moving-window variance and squared-difference," because "anomalies are not consistently harder to reconstruct or forecast." That is a finding about anomaly detection, not forecasting accuracy; the paper does not touch the leaderboards above, and reading it as "foundation models do not beat baselines" would be the overclaim this post exists to criticize. It lands anyway, because the skill Google's repo tells you to copy into your agent ships an anomaly-detection example, examples/anomaly-detection/detect_anomalies.py, 17,032 bytes of it, and that file loads the TimesFM 1.0 checkpoint, three generations behind the model on the leaderboards. The same model family is the best forecaster on three boards and, used the way that file uses it, no better than two lines of arithmetic. Nothing in the install path tells you which task, or which generation, you are holding.

GIFT-Eval tracks that as a column, declared by each submitter, and to Google's credit TimesFM-3's row reads leak: no, replication code: yes. All twelve systems ranked above it are typed agentic, and ten of those twelve publish no replication code at all. The most reproducible entry on the board sits below a dozen systems nobody outside their authors can re-run. That is the number to carry into a vendor meeting: not the rank, but whether anyone can reproduce it.

#The lesson underneath the model

The catch that shaped how we read this repo happened by accident, and it is worth stating plainly because it is the same failure this piece is warning about. One benchmark aggregator publishes a summary card for GIFT-Eval that lists a top three with no TimesFM entry at all, on a narrower slice of the same board. Read alone, that card falsifies Google's claim. The same site's detail page, sorted the same way, shows all 130 models and puts TimesFM-3 thirteenth overall, first among the non-agentic entries. Two views of one benchmark site, two different populations and metrics, opposite conclusions, and the one that looks like a "gotcha" is the one that stops at the summary card instead of opening the underlying table.

That is the same discipline our own buyer's guide opens with: ask whether the headline number is the actual result, or a number one step over from the table that would tell you otherwise. Google's TimesFM did not need a single invented statistic to make this post. Every number here is true, on its own terms, in its own document. The story is that no single page tells you which model, license, and behavior you are actually getting, and the two most commercially relevant pages, BigQuery and the agent skill, never once mention the other's terms.