Mistral released Timestral on October 15, a family of four pretrained time-series forecasting models at 5M, 28M, 123M and 209M parameters that share one architecture and one training recipe. Mistral’s announcement The pitch is zero-shot forecasting on ordinary hardware: Mistral says the 5M model produces a day-ahead forecast in 10 milliseconds on an 8-thread server CPU, and the 209M model in 140. Mistral’s announcement The weights are ungated on Hugging Face under a Modified MIT license that stops applying once your company’s global consolidated monthly revenue passes $20 million. LICENSE file
S5 Labs has not run any of the four models. Every benchmark and latency figure below is Mistral’s.
What Mistral shipped
Each model takes a context of up to 32,768 steps, handled as 1,024 patches of 32, and forecasts up to 1,024 steps in a single forward pass; longer horizons are produced block by block. Model card The 209M card lists 24 layers and an embedding dimension of 768. Model card Mistral’s docs page labels the 209M model a Public Preview with 209M total and active parameters, a 32K context, past-only and known-future covariates, and quantile outputs. Mistral docs
| Model | Parameters (safetensors) | Downloads, Oct 23 |
|---|---|---|
| Timestral-5M | 5,208,960 | 3,924 |
| Timestral-28M | 27,821,184 | 4,543 |
| Timestral-123M | 122,606,720 | 2,401 |
| Timestral-209M | 209,312,256 | 6,815 |
Source: Hugging Face API, checked October 23. The 28M repository was last modified on October 19, four days after the other three, for a reason Mistral has not stated.
Training was mostly synthetic. Mistral pretrained at a context of 8,192 on purely synthetic data, then post-trained on a mix of real and synthetic series at a context four times longer, with a pinball loss over 21 quantiles. Mistral’s announcement The blog does not say which real datasets went into post-training, and the Hugging Face card says a technical report is “coming soon”. Model card Mistral positions the models for on-prem, VPC or air-gapped deployment, with fine-tuning available through its Forge service. Mistral’s announcement
Installation is a pip install from GitHub, with Python 3.10 or later and PyTorch 2.4 or later; the runtime is PyTorch on CPU or GPU. Model card
from timestral import TimestralForecaster
model = TimestralForecaster.from_pretrained("mistralai/Timestral-209M", device="cpu")
forecast = model.forecast(context, prediction_length=96)
forecast.median
forecast.quantile(0.95)Which size to pick
Mistral’s own tables make the choice fairly mechanical. The four models land in order on every benchmark, and the gap between 123M and 209M is small everywhere.
| Benchmark (Mistral’s runs, zero-shot) | 5M | 28M | 123M | 209M |
|---|---|---|---|---|
| GIFT-Eval MASE (lower is better) | 0.705 | 0.678 | 0.659 | 0.658 |
| GIFT-Eval CRPS (lower is better) | 0.479 | 0.461 | 0.448 | 0.446 |
| TIME MASE | 0.652 | 0.638 | 0.628 | 0.626 |
| TIME CRPS | 0.545 | 0.534 | 0.525 | 0.522 |
| fev-bench SQL skill (higher is better) | 46.19 | 48.42 | 50.18 | 50.39 |
| fev-bench MASE skill | 34.54 | 37.27 | 39.19 | 39.49 |
Source: model card.
The latency figures come from one sentence in the blog: on “a standard server CPU with 8 threads”, forecasting one day ahead from three weeks of 15-minute history, the 5M model takes 10 ms and the 209M takes 140 ms. Mistral adds that the 28M is about five times faster than the 209M, and that the 123M comes within 0.2% of the 209M’s score at about three-quarters of its latency. Mistral’s announcement Working that through gives roughly 28 ms for the 28M and 105 ms for the 123M; those two numbers are our arithmetic, not Mistral’s. The blog does not name the CPU, the precision, the batch size or the model load time, and the model card gives no latency at all.
Read that way:
- 28M is the default. It is the smallest model that beats Chronos-2 in Mistral’s tables (more on that below), and at roughly 28 ms per forecast it is cheap enough to re-score every series on each refresh.
- 5M is for when the forecast runs inside a request path or on a device where 10 ms matters. It gives up measurable accuracy for that: 0.705 against 0.658 on GIFT-Eval MASE.
- 209M is for when accuracy is the only thing you are paying for and 140 ms per series is fine. For batch jobs that run overnight, the extra 35 ms or so over the 123M is nothing.
- 123M is a narrow middle. Within 0.2% of the 209M at three-quarters of its latency is a real saving only if you are latency-bound and already past the 28M’s accuracy.
Does it replace Chronos-2 or a tuned statistical baseline?
Against other pretrained forecasters, Mistral’s numbers favor the 209M on every row it reports. The comparison is zero-shot, Mistral-run, and the column models are the ones Mistral chose.
| Benchmark (Mistral’s runs) | Timestral-209M | t0-gamma | TimesFM-3 | Toto-2.0-2.5B | Chronos-2 |
|---|---|---|---|---|---|
| GIFT-Eval MASE | 0.658 | 0.661 | 0.667 | 0.696 | 0.698 |
| GIFT-Eval CRPS | 0.446 | 0.452 | 0.456 | 0.476 | 0.485 |
| TIME MASE | 0.626 | 0.637 | 0.640 | 0.642 | 0.662 |
| TIME CRPS | 0.522 | 0.534 | 0.536 | 0.539 | 0.556 |
| fev-bench SQL skill | 50.39 | 49.42 | 48.99 | 44.42 | 47.28 |
| fev-bench MASE skill | 39.49 | 37.99 | 37.82 | 32.54 | 35.50 |
Source: model card.
Mistral says the 209M ranks first among pretrained zero-shot models “across all five metrics” from a single checkpoint without ensembling, that it beats the previous state of the art with about two-thirds of that model’s parameters, and that it beats TimesFM-3 and Falcon-2.0 with fewer parameters. Mistral’s announcement The blog does not name the previous state of the art or give sizes for TimesFM-3 or Falcon-2.0. Its leaderboard charts show the top ten of 83 entries on GIFT-Eval, 35 on TIME and 37 on fev-bench, and Mistral says it has submitted the results to the GIFT-Eval and TIME leaderboards. Mistral’s announcement As of this writing we have no independent confirmation of any score or rank.
On Chronos-2 specifically: the 28M beats Chronos-2’s reported MASE on both GIFT-Eval (0.678 against 0.698) and TIME (0.638 against 0.662), and Mistral says Chronos-2 is about four times larger than the 28M, without stating which Chronos-2 variant it means. Mistral’s announcement The 5M is a mixed case: it edges Chronos-2 on TIME and on GIFT-Eval CRPS, but loses on GIFT-Eval MASE (0.705 against 0.698) and on both fev-bench skill scores. If you are on Chronos-2 today and the benchmark mix resembles your data, the 28M is the candidate to test; the 5M is not a drop-in replacement.
The statistical-baseline question has no answer in Mistral’s material. Neither the blog nor the card compares Timestral with a tuned ETS, ARIMA or seasonal-naive model on any dataset. For a team whose forecasts come from a well-maintained statistical pipeline, Timestral is a zero-shot model that may or may not beat what you already tuned; the only way to find out is to run both on your own holdout.
Quantile output for alerting and capacity planning
Every Timestral model returns 21 quantiles, from 0.01 to 0.99, and the Python API exposes them as .median and .quantile(q) on the forecast object. Model card A point forecast tells you the expected load; a 0.95 or 0.99 quantile tells you how much headroom to provision, and the gap between the 0.05 and 0.95 quantiles gives you an alert band that widens when the model is uncertain instead of a fixed threshold that fires on every quiet weekend. You still have to measure whether the bands are well calibrated on your data. Pinball loss is the standard quantile-regression objective, but Mistral has published no calibration results that we have seen.
The docs page and model card both confirm support for past-only covariates and known-future covariates, which covers the two cases most planning work needs: signals you observe alongside the target, and signals you know in advance such as holidays or scheduled promotions. Mistral docs Model card The model card also says covariates can be numeric or categorical, with categorical ones target-encoded, that missing values are marked as NaN, and that the models accept univariate and multivariate targets. Model card We have not tested any of that.
The license: MIT with a monthly revenue line
The LICENSE file in the 209M repository is an MIT-style grant over the weights and documentation with an attribution condition and one large exception: it grants no rights if the global consolidated monthly revenue of your company, or your employer, exceeded $20 million (or the equivalent) in the preceding month. LICENSE file The condition follows derivatives and combined works whether they come from Mistral or from a third party. Companies over the line are directed to [email protected] for a commercial license, granted at Mistral’s sole discretion, or to Mistral’s hosted services. The weights come as-is with no warranty, and the model card adds that the models must not be used in ways that infringe third-party rights. Model card Hugging Face tags the repository license:other. Model card In our reading, the revenue test means this is not open source by the OSI definition, however permissive it is below the line.
The threshold is monthly and tested one month at a time, so it sits well above the $10 million annual line Liquid draws for LFM2.5-24B-A2B and its d1 decision model. The clause covers your employer as well as your company, so a consultant deploying Timestral for a large client should read it with the client’s revenue in mind. We checked the LICENSE file in the 209M repository only; the other three sizes presumably carry the same text, but we did not open them.
What to test and what is still unknown
- Run the 28M and 209M on your own holdout against whatever you forecast with today, statistical or otherwise. Mistral’s tables contain no classical baseline, so that comparison is yours to make.
- Measure latency on your CPU at your series length and batch size. The 10 ms and 140 ms figures come from one unspecified machine on one workload.
- Check quantile calibration before wiring the 0.95 quantile into an autoscaler or an alert.
- Test the covariate behaviour you need, in particular target-encoded categorical covariates and NaN handling, which we have only read about in the model card.
- Have someone confirm which side of the $20 million monthly line your company, and your employer, sit on before an evaluation turns into a dependency.
Still missing from Mistral: the technical report, the real-data sources used in post-training, independent leaderboard confirmation, the Chronos-2 variant it compared against, hosted pricing, and the reason the 28M repository changed on October 19. Timestral is the second small, task-specific open-weight release we have covered this month, after IBM’s Granite Switch 4.1, and our open-source AI models guide tracks the licenses across them. For a team under the revenue line with forecasting work that fits on a CPU, the 28M is worth an afternoon of testing.