That makes license terms and per-dataset behaviour the real basis for choosing.
Summary
Mistral released Timestral, an open-weight family of time-series foundation models in 5M, 28M, 123M and 209M sizes. For the 209M model it reports a GIFT-Eval MASE of 0.658. IBM Research published Granite Timeseries Ensemble R1, which pools four Granite checkpoints. IBM reports a MASE of 0.683 on 97 GIFT-Eval configurations, against 0.685 for its strongest single member, and that member beats the ensemble on CRPS/MWQL. Neither vendor's figures have been independently reproduced in the material we have.
What happened
Timestral is Mistral's first forecasting model family. The 209M checkpoint on Hugging Face accepts up to 32,768 historical steps and forecasts up to 1,024 steps in one forward pass. It supports univariate and multivariate targets, past-only and known-future covariates, and 21 output quantiles. It runs in PyTorch on CPU or GPU. Mistral's product page lists on-prem, VPC and air-gapped deployment. The Hugging Face page says no inference provider hosts the model and lists the tech report as "coming soon". Mistral reports results on GIFT-Eval, TIME and fev-bench, including 0.658 MASE on GIFT-Eval and 0.626 on TIME. The weights use a Modified MIT license. The word "Modified" matters, because the license is not plain MIT.
IBM's Granite Timeseries Ensemble R1 is a configuration loaded through granite-tsfm. At runtime it downloads Granite PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1 and TTM-r3. The default Probabilistic Ensemble pools the members' quantile forecasts with equal weight and needs no retraining. An optional IQR Ensemble reweights members by the spread of their forecasts. On 97 GIFT-Eval configurations, IBM reports the default ensemble at MASE 0.683 and PatchTST-FM-r2 alone at 0.685. On CRPS/MWQL the single model is better, 0.467 against 0.471 for each ensemble variant. IBM describes the default recipe as permissively licensed but warns that expanded recipes can carry non-commercial or research-use terms.
What the archive adds
The archive shows this category filling up. NVIDIA released Kumo Forecast, an Apache-2.0 forecasting model and library that combines pretrained time-series representations with retrieval of historical patterns (d1). IBM's Granite time-series line has long been built for cheap deployment: models with a few million parameters and GPU-free inference (d2). The GIFT-Eval paper explains why every release now quotes the same scoreboard (d3). Before it, there was no unified, diverse zero-shot benchmark, so cross-model comparisons were hard. The same paper treats data leakage as a field-level risk. That warning matters for any pretrained forecaster whose training corpus is not fully published. A vendor statement that its pretraining data does not overlap the test series can only be checked as far as the vendor discloses that data.
Trend signal
Open forecasters are following the open-LLM pattern. Each lab ships a model family, each model card leads with one benchmark aggregate, and the reported gaps keep shrinking. Mistral, IBM and NVIDIA now all have open time-series models (primary evidence, d1, d2). They also compete on the same deployment pitch: small parameter counts, CPU inference, and no need to send operational data to a hosted endpoint. When several vendors sit a few hundredths of MASE apart on one aggregate, that aggregate stops working as a way to choose. What separates the models is how the score was produced (a shared harness or self-reported), the license terms, covariate support, context length, calibration, and per-dataset behaviour.
Why now
1) A shared scoreboard exists. GIFT-Eval gave the field a unified zero-shot benchmark where none existed (d3). That is what lets vendors headline a single MASE number, and it is also why small gaps between those numbers get overread. 2) Competition is shifting to deployment cost. IBM's Granite line targets few-million-parameter, GPU-free inference (d2), and Mistral advertises CPU, on-prem and air-gapped use (primary evidence). Both are pitching to enterprises that won't send operational data to an API. 3) Peers are filling the category. NVIDIA's Apache-2.0 Kumo Forecast (d1) and IBM's existing line made open forecasters something a lab like Mistral needs to offer.
Connections to other events
Here is what the numbers support and what they don't.
1) Both scores beat the naive baseline. MASE divides a model's error by the error of a seasonal-naive forecast, so anything below 1 beats that baseline. Taken at face value, both 0.658 and 0.683 show that zero-shot models beat naive forecasting across GIFT-Eval in aggregate without any tuning. That much holds up.
2) Timestral against Granite. 0.658 against 0.683 is a gap of 0.025, about 3.7% relative. That gap only means something if both numbers come from the same configuration set, aggregation and split. IBM states that its figure covers 97 configurations. The Timestral material we have does not state its configuration count or aggregation method, and the tech report is not out. So these two numbers cannot yet be compared like for like. The first party to publish per-configuration results through the public GIFT-Eval harness sets the reference point, and the other number has to be re-scored against it.
3) The ensemble against its best member. This comparison is internally consistent, because IBM measured both numbers on the same 97 configurations. The ensemble improves MASE from 0.685 to 0.683 and is worse on the probabilistic metric, 0.471 against 0.467. Our inference is that a 0.002 MASE gain averaged over heterogeneous configurations is probably smaller than the variation between datasets. IBM publishes no confidence intervals, so this cannot be confirmed. Either way, IBM's own table says that running four checkpoints gives roughly what PatchTST-FM-r2 gives alone, with slightly worse quantiles.
4) Cost is missing. Neither release reports latency, memory or cost per forecast. A 209M model and a few-million-parameter Granite member (d2) are very different choices when you score thousands of series every hour. An unverified 0.025 aggregate edge may not pay for that difference.
5) Contamination. GIFT-Eval's authors flag leakage (d3). Any claim of no train/test overlap is self-reported until the pretraining data is disclosed.
Why it matters
For builders, a zero-shot forecaster is now an afternoon baseline, and every serious option clears seasonal-naive in aggregate. The decision rests on other things. First, license: Timestral ships under a Modified MIT license, so commercial users should read the modifications in full before treating it as unrestricted. IBM's default Granite recipe is described as permissive, but expanded recipes may be non-commercial. Second, inputs: Timestral's 32k context and known-future covariates matter for long seasonality and for calendar- or promotion-driven demand. Third, cost: Granite's small members matter if you score huge numbers of series on CPU. Fourth, ensembling: IBM's own results give little reason to pay for four inference passes by default. Buyers should ask vendors for per-dataset results, CRPS and interval coverage in addition to aggregate MASE, and should ask whether the numbers came from the public harness.
What happens next
- Mistral publishes the Timestral tech report with per-dataset or per-configuration GIFT-Eval tables. (3 months, 55%). Watch for: The 'coming soon' tech-report link on the Hugging Face model card resolves to a document with per-dataset tables. Reasoning: the link is already promised publicly, but labs often slip report dates, so this is only slightly better than a coin flip.
- An independent group publishes a Timestral-209M GIFT-Eval MASE within 0.02 of Mistral's reported 0.658. (6 months, 50%). Watch for: A third-party leaderboard entry, paper or blog reporting Timestral-209M MASE between 0.638 and 0.678. Reasoning: open weights make reproduction easy, but differences in harness setup often shift aggregates by more than 0.02.
- IBM changes the default Granite ensemble recipe (different members or non-equal weighting), or recommends a single model as the default. (12 months, 30%). Watch for: A granite-tsfm release or model card that drops equal-weight pooling as the default. Reasoning: IBM's own CRPS figures favour the single member, but vendors rarely walk back a recipe they just launched, so the odds are low.
Limits and uncertainty
From the material we have, we could not confirm whether Timestral's GIFT-Eval results have been submitted to the official leaderboard with replication code, or exactly which configurations Mistral scored. If they match IBM's 97, the 0.025 gap counts for more than this piece allows. We also have not seen the full text of the Modified MIT license. Any commercial thresholds or carve-outs in it should be read directly from the license file, not from secondhand summaries. The claim that the ensemble's 0.002 gain is within noise is our inference, since no variance figures are published. Aggregates can also hide a decisive win in one domain such as energy or retail. No source measures CPU speed, so we make no speed claims.
Skill to build now
Learn paired comparison of forecasters instead of comparing aggregate MASE numbers. First exercise: 1) Pick 50 to 200 of your own series and three rolling forecast origins. 2) Score Timestral-28M and PatchTST-FM-r2 with MASE and CRPS for each series and origin. 3) Take the per-series differences and bootstrap them, resampling series 1,000 times, to get a 90% interval on the mean difference. 4) Repeat with GIFT-Eval per-dataset results wherever a vendor publishes them. If the interval includes zero, you have measured directly that the headline gap doesn't settle the choice.
Reader takeaway
Before adopting Timestral or Granite, read the Timestral license file for commercial restrictions. Then run Timestral-28M, PatchTST-FM-r2 and seasonal-naive on your own series over at least three rolling origins. Adopt a model only if a paired bootstrap shows it beating the alternative with an interval that excludes zero. If it doesn't, the cheaper and less restricted option wins.
Open question
When open forecasters from several labs land within a few hundredths of MASE of each other on GIFT-Eval, are the remaining differences forecasting skill, or mostly differences in harness choices and in how much each vendor's undisclosed pretraining data overlaps the test series?