Mistral's Timestral: the numbers hold, the first place does not
Mistral published four time-series models in mid-October 2026: Timestral with 5, 28, 123 and 209 million parameters, built for forecasting without domain knowledge. The model card and the paper submitted later claim top places on three public boards. We started recounting the values in October and measured all three boards again on 20 November 2026. In one sentence: every number Mistral reported is correct. The claim that the model sits at first place only holds if you read the mirror rather than the source board.
What is on disk, in bytes
Timestral-209M occupies 837.27 megabytes in a single weight file on Hugging Face, 209.3 million parameters; the small variant is 20.84 megabytes. The configuration file names a patch size of 32, 24 transformer layers, model width 768, 21 quantile knots, a context of 32,768 steps and 1,024 steps per pass. Installation goes through Git: pip install git+https://github.com/mistralai/timestral.git, Python 3.10 or newer, Torch 2.4 or newer, and there is no package on PyPI (GitHub issue #5).
Provenance is the strangest part. The repository mistralai/timestral has exactly one commit, dated 15 October 2026 at 14:59:55 UTC, and nothing has arrived since. There is no release and no version tag. On the measurement day the repository shows 1,231 stars, 103 forks and 9 open issues; the model card shows 30,653 downloads and 499 likes. All six evaluation entries on the card carry verified: false, which is routine there, but combined with the missing release it means nothing outside that single upload has been confirmed.
The scores: reproduced exactly
We did not quote the aggregation, we rebuilt it with the boards' own code. For GIFT-Eval that means: every task cell divided by the Seasonal Naive cell, then the geometric mean over tasks. For fev-bench the official path from the analysis module, including leakage imputation and clipping. The agreement is complete.
| Model | Card (gMASE / gWQL) | Our aggregation | fev-bench reported / recounted |
|---|---|---|---|
| Timestral-209M | 0.658 / 0.446 | 0.6575 / 0.4460 | 50.4 / 50.39 |
| Timestral-123M | not stated | 0.6589 / 0.4476 | not stated / 50.18 |
| t0-gamma (open, no licence cap) | not stated | 0.6608 / 0.4520 | not stated / 49.42 |
At 209 million parameters that is not a rounding story but a clean template: whoever reported the number aggregated it from the same files that are public. Which is exactly why the second half of the claim deserves a look, the ranking.
The ranking: three surfaces, three answers
The claim comes from a GitHub issue dated 16 October 2026: Timestral-209M results were said to be "currently first on two metrics" on GIFT-Eval, TIME and fev-bench. Mistral was asked to confirm; the question was never answered. On 20 November 2026 we measured the three places where that ranking could live today, separately.
| Surface | State today, measured | How we measured it |
|---|---|---|
| fev-bench source board | Timestral nowhere, the board lists 26 models | leaderboard file of the publication pipeline, 546,970 bytes, timestamp 20 Nov 2026 06:11 UTC; positive control from the model list inside the same file |
| GIFT-Eval source board | 174 model directories, but the public summary table names 21 models without Timestral | directory tree of the Space against the five aggregated CSV files |
| TIME output repository | 48 model directories, all four Timestral sizes, 481 individual files for 209M | result tree; we deliberately did not rebuild the official summary table |
| Papers with Code mirror | badges: 2nd on GIFT-Eval, "result retired" on TIME, "SOTA" on fev-bench | badge graphic per evaluation, four requests |
| Papers with Code page header | "Best: 0.658 by Timestral-209M" (GIFT), "Best: 0.626" (TIME), "Best: 50.4" (fev-bench) | delivered page, server-rendered line |
Two things stand out. First, the fev-bench source board does not contain Timestral at all, while the mirror credits the model with the best value there. Second, the mirror contradicts itself on the same day: the badge graphic reports second place on GIFT-Eval and a retired entry on TIME, while the header of that very evaluation still names Timestral as best. We cannot derive the rule behind it from the files; we only know the two readings. What we take from it is a measurement rule for our own boards: a ranking is a property of a surface on a day; the model itself holds no rank. A rank without board state, model count and timestamp carries no information.
Our own recount on the source tables (state of 23 October 2026) is the only number base we name for an ordering. On fev-bench: fourth of 43 models, and fourth on MASE skill as well with 39.49. On GIFT-Eval: 14th across all entries and second among open zero-shot models. TimeH2O-v1 leads that group with 0.6536 gMASE ahead of Timestral-209M at 0.6575, with t0-gamma behind at 0.6608. Ahead of all three sit six agentic entries and the closed Morpheus-2.0 at 0.6459. The gap to t0-gamma is 0.5 percent on MASE and 1.3 percent on WQL, which is the range where a different task mix flips the order.
The licence sits where nobody looks
On Hugging Face the card is marked as licence type other. The licence file itself is a modified MIT licence with one condition: it does not apply to companies that booked more than 20 million US dollars in revenue in a month. That condition lives in the licence file; the model card's README only shows the intellectual property line. The GitHub repository points to an unmodified Apache-2.0 licence, and GitHub reports Apache-2.0 as the repository's detected licence as well. Two primary sources, two statements, no clarification from the vendor. Anyone putting Timestral into a commercial pipeline has to settle that question before evaluating, and with t0-gamma there is an alternative with no cap that sits only 0.5 to 1.3 percent behind in our measurement.
The paper names one thing that decides it for us
On 27 October 2026, twelve days after the upload, a paper for Timestral appeared on arXiv (2610.37585, AI category, 356 authors, version v1 only). Neither the model card nor the README links it; the corresponding Hugging Face discussion, "Add link to the paper page", has been open since 29 October. We take two points from the abstract. The method is called Decoupled Horizon Forecasting and treats known future covariates as an input, which is the right approach for operations forecasting. And the model trains in two stages, the first one exclusively on a synthetic corpus, with real series only entering in the mixed post-training stage.
For us that is the deciding point, independent of any ranking. We forecast utilisation and freight costs, so series whose peaks are tied to holidays, toll-system changes and shift schedules. Exactly those covariates have no counterpart in a synthetic corpus. There is also a trap in the aggregation path that we read from the code: with symmetric_averaging=True and a covariate without dispersion (a holiday that occurs once in the window, a toll class with a single value), the symmetrization inflates the upper quantiles noticeably. Anyone using upper quantiles for capacity decisions has to check the window first. That warning stands in the source text; we have not measured it, see the limits below. That is why every candidate model has to clear our own series first, the way we did with the Jev decision model.
What we did not measure
So the limits are heard as well: we did not run a Timestral model in our stack. There are no throughput, latency or memory figures from us for this model, and we deliberately do not compare token counts with other models. We never read the weights in full, the download of the 837 megabytes stopped at 214 megabytes. We deliberately did not rebuild the TIME summary table; it is built from window-level metric files by a 43 kilobyte evaluation script, and a recount without a logged procedure would be a second claim rather than a proof. What we measured are values, rankings, licences, file inventories and the statements of the surfaces.
Assessment
Timestral is a cleanly documented, small time-series model whose numbers hold. We keep it as a candidate for operations forecasting, but as a watch item rather than a pilot: the licence question is open, the repository has not moved since one commit, the PyPI package is missing, and the synthetic-only pretraining hits exactly the weakest part of our data. The three reasons to re-enter are a second commit with a PyPI package, a clarified licence line, and a run in our own house, on top of our own AI stack, with real holiday and toll windows.
The rule we take from this measurement for our own work, and for any forwarded screenshot with a ranking, stays simple: recount the numbers, measure the rank where it is produced, and write the state of the list next to the value. The number 0.658 is checkable. First place was not, on the day of publication, and the two lines about it sit in the October folder of our dossiers.
Further reading
Five sources, all fetched by us on 20 November 2026; our own aggregation dates from 19 and 25 October 2026:
- Timestral-209M model card on Hugging Face (scores, configuration, licence file): https://huggingface.co/mistralai/Timestral-209M
- The
mistralai/timestralrepository on GitHub (one commit, the PyPI discussion in issue #5): https://github.com/mistralai/timestral - Version v1 of the paper on arXiv, submitted 27 October 2026: https://arxiv.org/abs/2610.37585
- fev-bench source board in the AutoGluon group's Streamlit space, whose board file feeds the publication pipeline: https://huggingface.co/spaces/autogluon/fev-bench
- The GIFT-Eval space with the result tree of every model: https://huggingface.co/spaces/Salesforce/GIFT-Eval
Is Timestral the best time-series model?+
Not according to the source boards. On fev-bench our own aggregation puts Timestral-209M fourth of 43 models, behind TimeRSI-1.5, TimeRSI and TimeH2O-v1. On GIFT-Eval it is 14th across all entries and second among the open zero-shot models, behind TimeH2O-v1.
Are the numbers Mistral reported correct?+
Yes. Our own aggregation from the task-level result files gives Timestral-209M 0.6575 gMASE and 0.4460 gWQL on GIFT-Eval and 50.39 skill score on fev-bench. The model card states 0.658, 0.446 and 50.4. The scores are right; only the ranking attached to them is not.
May we use Timestral commercially?+
The licence file on Hugging Face is a modified MIT licence with a condition: it does not apply to companies with more than 20 million US dollars in consolidated monthly revenue. That condition appears in the licence file, while the model card's README carries only the intellectual-property line. The GitHub repository shows an unmodified Apache-2.0 licence instead. Settle that discrepancy before any commercial deployment.
senn-tech