Timestral is a family of time-series forecasting models from Mistral, announced on 15 October 2026. It takes numeric series and returns probabilistic forecasts as 21 quantiles, not text. Mistral says it needs no GPU for inference.
The short answer
- Sizes: 5M, 28M, 123M and 209M parameters, with one architecture and training recipe. The fp32 weights are 21MB, 111MB, 490MB and 837MB.
- Output: probabilistic, with 21 quantiles from 0.01 to 0.99. It is multivariate, so it models related series jointly.
- Context and horizon: up to 32,768 context steps and a 1,024-step horizon in one pass.
- Hardware: Mistral says it “does not need a GPU at inference.” It needs Python 3.10+ and torch 2.4+, and it also runs on CUDA.
- License: code under Apache 2.0, weights under Modified MIT. The weights license sits in a per-model LICENSE file that was not opened.
- Benchmark claim: “ranks first on all three of the field’s main public benchmarks” is Mistral’s claim. It could not be confirmed from the leaderboards, and no independent evaluation was found.
Why is Timestral a forecasting model and not a chatbot?
Mistral calls Timestral “a numerical model.” It can run inside existing monitoring pipelines, or be paired with language models in agentic workflows. There, the language model handles context such as maintenance logs and reports, and Timestral handles the numeric forecast.
The model card calls it a “zero-shot forecasting foundation model” that takes numeric series and returns quantiles. It is not a substitute for asking a chat assistant to “predict sales.”
For longer horizons, the model forecasts block by block and feeds each block’s median back in.
What goes in and what comes out?
- Inputs: multivariate targets, past-only covariates, known-future covariates and categorical covariates. Missing values are NaN, with an optional observation mask.
- Outputs: forecast.quantiles shaped (series, variates, horizon, 21), for example (2,1,96,21) in the README, plus forecast.median.
- Covariate limit: covariates only condition the forecast, and only known-future covariates see the forecast horizon.
How fast is it on a CPU?
This is Mistral’s own measurement, not independently reproduced. On an 8-thread server CPU, a one-day-ahead forecast from three weeks of 15-minute history takes 10 ms for the 5M model and 140 ms for the 209M.
Mistral says the 123M model comes within 0.2% of the 209M on accuracy at about three-quarters of the latency. The 28M is about five times faster than the 209M. No memory or VRAM figures were found.
What does Mistral claim on benchmarks?
| Benchmark | Mistral’s claim |
|---|---|
| GIFT-Eval | First among pretrained, zero-shot models; 97 configs with default settings |
| TIME | First across 98 tasks |
| fev-bench | First across 100 tasks |
| 209M model card | GIFT-Eval MASE 0.658 and CRPS 0.446 TIME MASE 0.626 and CRPS 0.522 fev-bench SQL skill score 50.39 and MASE skill score 39.49 |
The blog says the 209M model beats the previous state of the art with two-thirds of its parameters, from a single checkpoint with no ensembling. The README says all four sizes sit on the accuracy-versus-size Pareto front “at launch.” All of these are Mistral’s figures.
The GIFT-Eval and fev-bench Spaces loaded without ranking rows, so Timestral’s position could not be confirmed. The TIME leaderboard was not opened.
A search result that was not opened mentions a different model reported to lead TIME and the GIFT-Eval pretrained category. That could conflict with “first on all three,” so check the leaderboards directly.
What limits does Mistral state?
- The full technical report is “coming soon.”
- Context is capped at 32,768 steps, and long horizons use block-wise forecasting.
- Support is “best-effort,” and the repository does not accept pull requests.
- The card and README have no limitations section. That public benchmarks may not match your production data is an inference, not Mistral’s statement.
How can you try it?
- Pick a numeric series with enough history (demand, load, a sensor) and a clear metric, such as absolute error against a simple baseline.
- Compare it with a conventional forecaster on the same test window, not with a chat assistant.
- Use the quantiles, not just the median, when the decision is sensitive to risk.
For the difference between a model that makes a structured decision and one that generates text, see Microsoft Decision-1: a decision model, not a generative model.
Questions and answers
Is Timestral a chat model?
No. It takes numeric series and outputs probabilistic quantiles.
Does it need a GPU?
Not for inference, per Mistral. It runs on CPU and can use CUDA.
Is it the best forecasting model?
Mistral claims first place on three benchmarks. That could not be confirmed from the leaderboards.
What is the license?
Code is Apache 2.0 and weights are Modified MIT.
Terms in plain language
- Time series: numbers recorded over time, such as daily demand or sensor readings.
- Quantile: a point in the range of possible outcomes. Several quantiles together show how uncertain a forecast is.
- Zero-shot: the model forecasts a new series without being retrained on it.
- Covariate: an extra series that helps explain the target, such as weather.
- Inference: running a trained model to get results.
Sources
Facts last checked against the linked pages on 18 October 2026. Benchmark figures are Mistral’s.
