Europe is competing at the front of forecasting AI
The two largest Timestral models lead a published comparison of time-series foundation models across three benchmark suites. Explore the scores and see how TiRex-2 and international models compare.
Top 10 in the published comparison
Blue = European team or research base highlighted here. Gray = other / not classified here. Bar length starts at zero. Entries include different checkpoints of the same family.
The European story
Timestral-209M leads the reported comparison. The 123M checkpoint follows closely, ahead of the compared models outside its family.
t0-gamma is a leading non-Timestral competitor; its team reports research activity in Paris.
TiRex-2 is another European contender. GIFT-Eval lists separate pretrained (PT) and zero-shot (ZS) submissions; fev-bench uses a benchmark-specific checkpoint. They should not be treated as identical runs.
Method, sources and limits
Scores are copied from Table 3 of the Timestral technical report. Choose a metric to rank the reported entries with a score for it. Entries without a score for that metric are excluded; the displayed top ten is the top ten within this table. The scores are from the report's October 15 release-date snapshot, not a current or independently verified leaderboard.
GIFT-Eval and TIME show geometric-mean errors relative to seasonal naive: below 1 means lower aggregate error on that benchmark, not a guarantee on each series. fev-bench reports scaled quantile-loss skill in percent, where higher is better. Do not compare numbers across different metrics. Small differences in aggregated results need not be meaningful on your business data.
Public benchmark entries can differ in inputs, checkpoints and evaluation settings. On fev-bench, TiRex-2 uses a benchmark-specific checkpoint. Mistral says Timestral leads among pretrained zero-shot models, but these views include all rows in its report with available values, including that TiRex-2 entry. We do not assert that every checkpoint in the comparison has identical data access or training.
The live GIFT-Eval leaderboard also includes agents and other systems and uses a separate ranking measure. It can show a different order. Check each model’s license, test on your own data against relevant baselines, and consider uncertainty calibration and operational cost before deployment.
Additional context: NXAI TiRex-2 · t0-gamma release · Forecasting Company Paris research.