Time Series Foundation Models: Timestral Tops GIFT-Eval, but Its License Ends at $20M a Month

Mistral released Timestral on 15 October. It is a family of four time series foundation models at 5M, 28M, 123M and 209M parameters, with probabilistic, multivariate forecasts, past-only and known-future covariates, and a 32,768-step context. The announcement says the 209M model ranks first on GIFT-Eval, TIME and fev-bench, runs on CPU in 140 ms for a one-day forecast from three weeks of 15-minute history, and can be deployed on-prem or air-gapped. The examples it gives are industrial plants with thousands of sensors and finance teams forecasting across thousands of accounts.

The benchmark claim holds up. The license is where a planning decision changes.

The license on the weights

The announcement and the model card both call the license Modified MIT. The LICENSE file is identical in all four Hugging Face repositories (same SHA-256), and the Hub tags it other. The modification is one condition. You may not exercise any rights under the license if the global consolidated monthly revenue of your company, or of your employer, exceeded $20 million in the preceding month. The condition covers derivatives, modifications and combined works, so fine-tuning the weights in-house does not get you out of it. Above the line, the license points you to Mistral's sales team for a commercial license, granted at its discretion, or to Mistral's hosted services.

$20 million a month is $240 million a year. The GitHub repository is a separate case: its inference code is Apache 2.0, and the README says the weights are not covered by that license.

The question was raised publicly. Issue 3 on the repository, opened 17 October, asks what the modification is, which real datasets went into post-training, and whether benchmark test windows were excluded. As of today it is open, and its only reply is a +1 from another user. The answer to the license question is in the LICENSE file. The training-data questions remain open until the promised technical report appears.

Does the time series foundation models leaderboard result reproduce?

Yes. GIFT-Eval's Space added the Timestral result files on 16 October, and the submission's config marks test-data leakage as No. We downloaded the per-configuration CSVs for Timestral, its rivals and Seasonal Naive. For each of the 97 configurations we scaled each model's MASE and quantile loss by Seasonal Naive's, then took the geometric mean. The result matches the model card to three decimals:

Lower is better. Over the whole benchmark, Timestral-209M has the lowest quantile loss on 65 of 97 configurations against t0-gamma and 81 of 97 against Chronos-2.

Where the ranking flips

The aggregate hides a domain split. Restricted to GIFT-Eval's Energy domain (32 configurations, the largest), t0-gamma scores 0.532 CRPS against Timestral-209M's 0.543, and 0.742 MASE against 0.772. It does this despite having lower loss on only 9 of the 32 configurations, so its wins are larger than its losses. In Econ/Fin (6 configurations) the two are level: 0.704 against 0.706. Timestral leads in Healthcare, Nature, Transport and Web/CloudOps, and narrowly in Sales.

The announcement names energy and finance among the decisions forecasts drive. Those are the two GIFT-Eval domains where Timestral does not lead. t0-gamma is not the smaller model: its checkpoint has 255.6M parameters to Timestral's 209.3M.

What to decide differently this quarter

The assumption this breaks is that open weights at the top of a forecasting leaderboard are a free default for an enterprise pilot. In this table, two of the top three entries carry restrictive licenses. TimesFM-3's license covers non-commercial, non-production use only. The model that wins on aggregate grants nothing to companies above the cap.

If your company is above the revenue line, take Timestral off the self-hosted shortlist. The alternative is a commercial negotiation, which is the buy side of the buy, build or fine-tune decision, not the build side. The derivatives clause removes the in-house fine-tuning route. Start the pilot on t0-gamma, with Chronos-2 as the second baseline. For series like GIFT-Eval's Energy set, the leaderboard says you are not giving up accuracy by doing so. This is the same lesson as Kumo Tabular: for a foundation model aimed at structured data, the license decides more than the leaderboard does.

If you are under the line, Timestral is a reasonable first candidate. Write down that the right lapses in any month your revenue crosses $20 million, because a growth year then becomes a migration project.

The check takes a minute:

curl -sL https://huggingface.co/mistralai/Timestral-209M/raw/main/LICENSE | grep -c million
curl -sL https://huggingface.co/theforecastingcompany/t0-gamma/raw/main/LICENSE | grep -c million

The first returns 1, the revenue clause. The second returns 0. Then backtest both models on your own series, scored only on windows after any plausible training cutoff, since issue 3 has no answer yet. Adopt the restricted model only if its quantile loss beats the open one by more than the 1.3 percent aggregate gap measured above. Below that margin, you would be taking on a licensing dependency for a smaller gain than the leaderboard shows. This is the kind of model-selection call our machine learning consulting work starts with.

What would make this wrong

Three things would change the recommendation. Mistral could publish a different license for the weights. The technical report could show that t0-gamma's Energy lead comes from training overlap with GIFT-Eval's test windows. Or your own post-cutoff backtest could show Timestral ahead by more than the threshold on your data. Any one of them moves Timestral back onto the shortlist.