Mistral has released a family of forecasting models that does not try to predict one inevitable future. Timestral-209M produces 21 quantiles—a distribution ranging from very low to very high outcomes—while accepting multiple related series and information that is genuinely known in advance. On the company’s published comparisons, the 209-million-parameter checkpoint ranks first across GIFT-Eval, TIME and fev-bench among the listed pretrained zero-shot models.[1][2]
The benchmark lead is the easy part to read. The harder deployment question is what information the model is allowed to see. A hospital can legitimately tell it that Monday is a public holiday or that an operating theatre is scheduled to open at 08:00. It cannot feed in the actual Monday admissions and then describe the result as a Sunday forecast. Support for future covariates makes the model more useful—and makes leakage controls more important.[1][2]
A forecast becomes a range rather than a single line
Timestral’s output quantiles can answer different operational questions. The median is a central forecast. The 10th and 90th percentiles can form a planning interval. A power operator might schedule ordinary generation around the median while reserving extra capacity for the high-demand tail. The model has not proved that those intervals are calibrated for every new dataset; the point is that its output preserves uncertainty instead of hiding it inside one number.[2]
What the 209M checkpoint accepts and returns
| Measure | Value | Context |
|---|---|---|
| Model size | 209 million parameters | The family also includes 5M, 28M and 123M checkpoints.Source 2 |
| Historical context | Up to 32,768 time steps | The architecture uses 32-step patches, giving up to 1,024 historical patches.Source 2 |
| Forecast horizon | Up to 1,024 time steps per forward pass | Longer horizons are produced block by block using the median forecast as new context.Source 2 |
| Probability output | 21 quantiles | The published set runs from 0.01 and 0.05 through 0.99.Source 2 |
| Inputs | Univariate or multivariate targets plus covariates | Past-only and known-future covariates may be numeric or categorical.Source 2 |
| Execution | PyTorch on CPU or GPU | The release is designed for local as well as accelerated inference.Source 2 |
Known-future information has a precise meaning
Which future variables are legitimate at prediction time
| Candidate input | Usually legitimate | Leakage risk |
|---|---|---|
| Calendar | Day of week, holidays, month-end or school term already known | Low, provided the calendar existed at the forecast dateSource 2 |
| Operations | Published opening hours, booked capacity or a planned maintenance window | Low to medium: use only the schedule that was actually availableSource 2 |
| Prices | A contractually fixed tariff or announced price change | Medium: do not substitute a later revised priceSource 2 |
| Weather | A forecast issued before the target period | Medium: archive the forecast vintage rather than using observed weatherSource 2 |
| Outcome itself | Actual demand, admissions, sales or failures during the period being predicted | High: this converts evaluation into hindsightSource 2 |
This distinction matters even in a careful backtest. A dataset can contain a weather forecast, a final weather observation and a later-corrected calendar in adjacent columns. Unless each row records the information vintage, a model may appear more accurate because the evaluation silently gave it cleaner knowledge than a live system would possess.
The benchmark lead is small in some places and larger in others
Published zero-shot results for Timestral-209M
| Benchmark | Metric | Timestral-209M | Next listed result |
|---|---|---|---|
| GIFT-Eval | MASE, lower is better | 0.658 | 0.661 for t0-gammaSource 2 |
| GIFT-Eval | CRPS, lower is better | 0.446 | 0.452 for t0-gammaSource 2 |
| TIME | MASE, lower is better | 0.626 | 0.637 for t0-gammaSource 2 |
| TIME | CRPS, lower is better | 0.522 | 0.534 for t0-gammaSource 2 |
| fev-bench | SQL skill score, higher is better | 50.39 | 49.42 for t0-gammaSource 2 |
| fev-bench | MASE skill score, higher is better | 39.49 | 37.99 for t0-gammaSource 2 |
The family comparison also shows diminishing returns. On GIFT-Eval MASE, the 123M checkpoint scores 0.659 and the 209M checkpoint 0.658. A deployment that values memory, CPU latency or many simultaneous series may rationally choose the smaller model. The largest model is not automatically the best economic choice.[2]
A responsible evaluation needs a time machine it does not have
The strongest test is a rolling historical simulation. At each cutoff date, freeze the exact data, forecasts, schedules and business rules available then; ask the model for a full probability distribution; and score both central error and calibration. Compare against a seasonal baseline and the current production method. That process is less glamorous than one leaderboard row, but it answers the question a deployment actually faces.
Timestral’s release is significant because it puts capable probabilistic forecasting into a relatively small, locally runnable model family. Its deeper contribution is to make forecast inputs explicit. The team that benefits most will not merely ask whether the model predicts tomorrow. It will prove that the model only knew what tomorrow had already revealed.
Sources & further reading
- Introducing Timestral Mistral AI · 15 Oct 2026 · Accessed 16 Oct 2026
- Timestral-209M model card Mistral AI / Hugging Face · Accessed 16 Oct 2026
- Timestral source repository Mistral AI / GitHub · Accessed 16 Oct 2026