Timestral tutorial: time series forecasting without training your own model

On this page
On October 15, 2026, Mistral AI released Timestral, a family of foundation models for time series. The largest model has 209 million parameters and, according to Mistral, ranked first on all three major public forecasting benchmarks at release. Code and weights are freely available, and the model runs on a GPU as well as on an ordinary CPU.
Until now, anyone who wanted to forecast electricity demand, store sales or a machine's temperature built a separate model for each task. An ARIMA or ETS model for single series, a gradient boosting model with dozens of handcrafted features for many series with external drivers, and yet another method for anomalies. Each of these models has to be trained, monitored and retrained. Timestral replaces this with a single pretrained model that you call with the history of a series and that returns a forecast with an uncertainty band right away.
In this tutorial we try that on real data: electricity demand and solar generation in Austria, daily sales of 1115 Rossmann stores, and a temperature sensor on an industrial machine. The electricity test lies entirely after Timestral's release, so the test data cannot have been part of training. We compare against classical methods, LightGBM and Amazon's Chronos-2, and also show where Timestral gets it wrong.
What is a foundation model for time series?
A language model has seen so much text during training that it can infer from a prompt how the text continues. A time series model like Timestral does the same with sequences of numbers. During training it has seen a large number of series with trends, daily and weekly patterns, jumps, gaps and noise. Given the history of your series, it recognizes these patterns and continues them. This learning from context is called zero-shot forecasting: you train nothing and only pass the history.
According to Mistral, most of the training data was synthetic. In the first stage, the model only saw generated series with a context of 8192 time steps. It was then trained for a shorter time on a mix of real and synthetic data with a four times longer context of 32,768 steps. Mistral has not yet published which real datasets were used. We come back to this in the section on limitations.
According to the model card, Timestral can:
- forecast one or several related series jointly (multivariate, for example vibration, temperature and pressure of a pump),
- take into account external drivers that are known only for the past or also for the future (covariates, numeric or categorical),
- read up to 32,768 time steps of history,
- predict up to 1024 steps in a single pass, and longer horizons block by block,
- output 21 quantiles per time step instead of a single value, from the 1% to the 99% quantile.
How Timestral reads a time series
How Timestral reads a time series: patches as tokens, forecast tokens for the horizon, and covariates that inform the target
Timestral is a transformer, the same kind of architecture as a language model. Instead of words, it processes segments of the time series:
- The series is split into pieces of 32 values each, called patches. Each piece is normalized first, and causally: mean and spread only come from values that lie before it. An asinh transform dampens outliers. Together with a marker for missing values, this becomes a token. 32,768 steps of context are therefore 1024 tokens.
- For the future, the model appends learned placeholders, the forecast tokens, one per patch of the horizon. Each placeholder reads the whole history before it through attention and outputs 21 quantiles for each of its 32 time steps. Mistral calls this Decoupled Horizon Forecasting. Unlike models that mask parts of the context, every observed value stays both context and target during training.
- Most of the transformer layers look along time, and each token only sees its past. Every fourth layer looks across the series. There, information flows from the covariates to the target, never the other way round. Known covariates such as a holiday calendar or a weather forecast may also look into the horizon. Past covariates such as a measured temperature end with the context.
- The output head produces the median and fans out the other quantiles from it so they never cross. Training used the pinball loss over all 21 quantiles.
Two tricks run automatically with every forecast. With symmetric averaging,
the model also forecasts the mirrored series and averages both results, which
costs two passes. And if the known history was never negative, the forecast is
clipped at 0. You can switch both off with symmetric_averaging=False and
clip_nonnegative=False.
The four sizes share the same architecture and training:
| Model | Parameters | Layers | Width | Weights (fp32) |
|---|---|---|---|---|
| Timestral-5M | 5.2M | 8 | 192 | 21 MB |
| Timestral-28M | 27.8M | 12 | 384 | 111 MB |
| Timestral-123M | 122.6M | 20 | 640 | 490 MB |
| Timestral-209M | 209.3M | 24 | 768 | 837 MB |
What the benchmarks say
Mistral measured Timestral on three benchmarks without any fine-tuning: GIFT-Eval (97 configurations), TIME (98 tasks) and fev-bench (100 tasks, many of them with covariates). The figures from the model card, as of release:
| Benchmark | Metric | Timestral-209M | t0-gamma | TimesFM-3 | Toto-2.0-2.5B | Chronos-2 |
|---|---|---|---|---|---|---|
| GIFT-Eval | MASE (lower is better) | 0.658 | 0.661 | 0.667 | 0.696 | 0.698 |
| GIFT-Eval | CRPS (lower is better) | 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 score (higher is better) | 50.39% | 49.42% | 48.99% | 44.42% | 47.28% |
MASE compares the error of the median with a naive seasonal forecast, CRPS scores the whole distribution. A value of 0.658 roughly means a third less error than "same as last season". The gaps to the next models are small. More remarkable is that, according to Mistral, even the 28M model beats Chronos-2 on GIFT-Eval and TIME at a quarter of its size.
Mistral measured these numbers itself, and it is unknown which real data went into training. On the GitHub page, a user asks exactly that, because public datasets could otherwise end up in both training and test. That is why we test below with electricity data from October 16 to November 12, 2026. It did not exist when the model was trained.
Installation
Timestral is not on PyPI yet. You install it directly from the GitHub repository and need Python 3.10 or later and PyTorch 2.4 or later:
For production environments, pin a commit, for example
git+https://github.com/mistralai/timestral.git@a534bdb. On first use,
from_pretrained downloads the weights from the Hugging Face Hub into the
local cache.
Step 1: Load the data
Our first example is Austria's electricity demand in 15-minute steps. The Energy-Charts API from Fraunhofer ISE provides consumption and generation for free and without a key. It does limit requests per minute, so the function waits briefly on status 429:
A little over a year of 15-minute values is about 39,000 time steps, more than Timestral can read. That is not a problem: the model automatically takes the last 32,768 steps, just under 341 days.
Step 2: The first forecast
The forecast itself is a single call. We use everything up to midnight (UTC) on November 5, 2026 as history and forecast the next 24 hours, that is 96 steps of 15 minutes:
MAPE 1.89 %
inside 80% interval: 77 %
Without training, without features and without a calendar, the forecast is off by 1.9% on average. 77% of the actual values fall into the band between the 10% and 90% quantiles. A well-calibrated model would get 80%.
About the shape of the data: forecast accepts a single series (T,),
several series (B, T) or several series with several variates (B, V, T).
The result always has the shape (B, V, horizon, 21). Missing values are
passed as NaN. The model does not see the timestamps themselves. It knows
neither weekday nor time of day and recognizes patterns only from the order of
the values.
How good is that in comparison?
A single day says little. So for each of the 28 days from October 16 to November 12, 2026, we created a forecast at midnight (UTC) for the next 24 hours and compared these methods:
- Seasonal naive: the value from one week earlier at the same time of day. Any serious forecast should beat it.
- MSTL from statsforecast decomposes the series into daily and weekly patterns plus trend, a classical statistical method, fitted on eight weeks of history.
- LightGBM is a gradient boosting model trained on three years of data before the test period. It gets the values from one, two and three days and one and two weeks earlier, yesterday's daily mean, quarter hour, weekday, day of year, holidays and temperature. This is what a typical model from a data science project looks like.
- Chronos-2 from Amazon (120M parameters) is a widely used time series model that also handles covariates.
| Method | MAPE | MAE |
|---|---|---|
| Seasonal naive (previous week) | 4.80% | 331 MW |
| MSTL (statsforecast) | 3.05% | 205 MW |
| Chronos-2 without covariates | 2.96% | 202 MW |
| Chronos-2 + day type | 2.36% | 159 MW |
| LightGBM (lags, calendar, holidays, temperature) | 2.08% | 142 MW |
| Timestral-209M without covariates | 2.35% | 160 MW |
| Timestral-28M + day type | 1.88% | 130 MW |
| Timestral-123M + day type | 1.76% | 121 MW |
| Timestral-209M + day type | 1.95% | 133 MW |
Without any extra information, Timestral clearly beats MSTL and Chronos-2. It only beats LightGBM once it gets a single additional piece of information, the day type, which the next section covers. 28 days is not a long test period, so we compared the daily errors pairwise: Timestral-123M with day type was better than LightGBM on 19 of 28 days, by 21 MW on average, and the 90% confidence interval of that gap, 5 to 37 MW, lies entirely below zero. For 28M and 209M the lead is smaller and not statistically significant.
Notably, the largest model is not the best one here. This matches Mistral's own numbers, where 123M and 209M are almost level. With a single series over four weeks, chance plays a part. So try several sizes on your data.
Step 3: Holidays as a covariate
Almost the entire gap of Timestral without covariates comes from one day. October 26 is Austria's national holiday, and demand looks like a Sunday. Timestral-209M saw a Monday coming and forecast a working day: a mean error of 944 MW instead of about 110 on normal days. LightGBM knows the "holiday" feature from three years of training and was at 165 MW on that day.
Electricity load in Austria from October 22 to November 2, 2026: without a calendar Timestral forecasts a working day on the national holiday, with the day type it gets almost as close as LightGBM
This is the most important property of a zero-shot model: it only knows what is in the context. A holiday appears only about a dozen times in 341 days of history, and the model sees no calendar. The solution is a known covariate that you pass for the past and for the forecast horizon.
We compared three variants. The first is a simple 0/1 "holiday" indicator. The second is a day type with three categories: working day, Saturday, Sunday or holiday. This trick is well known from classical load forecasting. A holiday is treated like a Sunday, and Sundays occur often enough in the context for the model to know what they look like.
With the 123M model, this gives for the holiday:
plain MAE 1149 MW
with day type MAE 247 MW
You pass categorical covariates as numeric codes. Timestral converts them internally with target encoding. Each category is replaced by the mean demand the series had in that category within the context. "Day type 2" thus becomes "on such days demand averaged about 5800 MW". The model can work with that number directly.
Across all 28 days, the covariates compare like this (MAPE):
| Covariates | Timestral-123M | Timestral-209M |
|---|---|---|
| none | 2.50% | 2.35% |
| holiday (0/1) + temperature | 2.12% | 2.42% |
| day type (categorical) | 1.76% | 1.95% |
| day type + temperature | 2.15% | 2.59% |
The second result surprised us. Temperature, a useful input for LightGBM (without it, its error rose from 2.1 to 2.6%), made Timestral worse in every combination. We suspect that the link between temperature and demand in Austria is weak in autumn and barely shows within a year of context, while LightGBM learned it from three years. We have not proven this. The practical rule is clear, though: measure each covariate on its own, more is not automatically better.
Step 4: Weather forecast for solar power
For solar power, on the other hand, weather is decisive. From the history alone, no model can know whether there will be fog tomorrow. Without covariates, Timestral is therefore no better than yesterday's value.
As a known covariate we use the forecast of global irradiance for five cities (Vienna, Linz, Graz, Salzburg, Innsbruck), averaged. It is important that you use a real weather forecast here and not the irradiance measured later. Otherwise you calculate with knowledge you did not have the day before, and the results look too good. Open-Meteo offers an archive of past weather model runs for this:
Across the 28 test days (error relative to mean generation, night hours included):
| Method | rel. MAE | MAE |
|---|---|---|
| Yesterday's value | 37.4% | 135 MW |
| Timestral-209M without weather | 37.1% | 133 MW |
| LightGBM (lags, calendar, irradiance forecast) | 23.4% | 84 MW |
| Chronos-2 + irradiance forecast | 19.8% | 71 MW |
| Timestral-5M + irradiance forecast | 20.8% | 75 MW |
| Timestral-28M + irradiance forecast | 18.4% | 66 MW |
| Timestral-123M + irradiance forecast | 16.2% | 58 MW |
| Timestral-209M + irradiance forecast | 17.3% | 62 MW |
Solar power in Austria from November 1 to 6, 2026: without weather, Timestral is too high on dull days and too low on sunny ones; with the irradiance forecast it follows the actual curve
With the irradiance forecast the error halves, and every Timestral size beats the LightGBM model, which had the same information and three years of training. Timestral-123M was better on 23 of the 28 days. The chart also shows the limit: on November 5, every method was too low at the midday peak, with Timestral closest at about 1600 instead of almost 2000 MW. The model cannot be more accurate than the weather forecast it gets.
Timestral's 80% interval was somewhat too narrow during daylight hours. For the 123M model 76% of values fell inside it, for the 209M model 68%. If you use the quantiles for decisions, check their coverage on your own data.
Step 5: A thousand series at once
Time series projects in retail rarely have one series, they have thousands: stores, products, regions. The Rossmann dataset from a 2015 Kaggle competition contains daily sales of 1115 stores of the German drugstore chain over 942 days, plus whether the store was open, whether a promotion ran, whether there were school holidays and whether it was a public holiday. The winners back then relied on extensive feature engineering and gradient boosting.
We forecast the last six weeks (42 days) for all stores. The data is available preprocessed on the Hugging Face Hub:
The chunks of 32 series have a reason. The variate attention runs over all rows of a call, that is series times (1 + covariates). Memory and compute therefore grow quadratically. A user reports on GitHub that a single call with 635 series and covariates tried to allocate 62 GB of memory. With 16 to 64 series per call it stays manageable.
About 180 stores have a gap of half a year in the data. The NaN values go
straight to the model, no preprocessing needed.
| Method | RMSPE | MASE | Time (GPU) |
|---|---|---|---|
| Seasonal naive (previous week) | 0.350 | 0.62 | - |
| LightGBM (store, calendar, promo, store profile) | 0.135 | 0.31 | 17 s training |
| Timestral-28M without covariates | 0.153 | 0.34 | 1.6 s |
| Timestral-28M + open, promo, school, holiday | 0.133 | 0.29 | 3.2 s |
| Timestral-209M without covariates | 0.126 | 0.27 | 1.8 s |
| Timestral-209M + open, promo, school, holiday | 0.129 | 0.28 | 16 s |
RMSPE was the metric of the Kaggle competition, computed over open days. For all methods we set days when the store was closed to 0. Our LightGBM is a solid baseline, not a competition solution. The winning solution reached about 0.10 on the competition's test period.
Timestral-209M beats the baseline without any covariate. Promotions in this data mostly run every other week, so the model can partly read them from the history. The small 28M model needs the covariates to catch up. For 209M they mainly improve the quantiles. Without covariates, the 90% interval contained 98% of the values and was too wide. With covariates it was 90%, exactly right.
Rossmann store 11: Timestral-209M gets the promotion weeks right, the band widens in the last promotion week
This is also where the quantiles become useful in practice. For staffing or order quantities, the median is rarely the right number. If you want to run short of stock or staff in only 10% of cases, plan with the 90% quantile:
A note on this test: Rossmann is part of fev-bench. We do not know whether the data was also in Timestral's training set. The electricity data above is therefore the cleaner test.
Step 6: Detecting anomalies
The quantiles can also be used to find unusual behavior. The idea: whatever falls outside the band the model considered plausible is suspicious. We test this on a temperature sensor of an industrial machine from the Numenta Anomaly Benchmark (NAB). The series has 22,683 values at five-minute intervals over a little more than ten weeks and four labeled incidents. According to the description, the first was a planned shutdown and another one a catastrophic failure of the machine.
The obvious approach is to forecast the next hour every hour and flag values outside the band. It did not work. The February incident starts with the temperature dropping from about 92 to 65 and then slowly sinks below 30 over a little more than a day. A model that forecasts anew every hour adapts to each new value and treats the slow decline as the normal course.
It works better if you compare the current measurement with an older expectation. We forecast 24 hours ahead every hour and only score the last hour of each forecast. Each measurement is thus judged against what the model expected a day earlier:
The roughly 1700 forecasts over 24 hours take a little over a second on an RTX 4080. How many incidents you find and how many false alarms you get depends heavily on two settings, the lead time and the width of the band:
| Lead time | Band | Incidents found (of 4) | False alarms in 10 weeks |
|---|---|---|---|
| 1 h | 1 to 99% | 0 | 0 |
| 1 h | 5 to 95% | 0 | 1 |
| 6 h | 1 to 99% | 2 | 7 |
| 6 h | 5 to 95% | 4 | 35 |
| 24 h | 1 to 99% | 2 | 0 |
| 24 h | 5 to 95% | 4 | 22 |
Temperature sensor of an industrial machine: alarms from the forecast band and the detail around the February 2014 failure
With a 24-hour lead time and the wide band, the method finds the two labeled incidents in December without a single false alarm. With the narrower band it finds all four, the February incident about eight hours after the first temperature drop, but also a little more than two false alarms per week. As a first filter that points an expert to unusual hours, this is useful. As the only alarm system it is not enough.
We tried the settings on exactly this series, so the table is on the optimistic side. For your sensors the same applies as for the threshold of any alarm system: you need a sample with known incidents to set lead time and band. In the technical report, Mistral names embeddings from the model for anomaly detection and classification as a next step. The published code does not include that yet.
Gaps and context length
Two questions come up in almost every project: what happens with missing data, and how much history do you need? We measured both on electricity load with day type (28 test days, MAPE):
| Context | Timestral-123M | Timestral-209M |
|---|---|---|
| 7 days | 2.12% | 2.02% |
| 28 days | 1.89% | 1.97% |
| 84 days | 2.23% | 2.27% |
| 341 days (maximum) | 1.76% | 1.95% |
| 341 days, 10% of values missing | 1.85% | 2.06% |
| 341 days, 30% of values missing | 1.97% | 2.23% |
Even one week of history gives a usable forecast. More context was usually better, but not always: 84 days were worse than 28. Check this on your own data as well instead of always using the maximum.
The model copes well with missing values. We put randomly placed gaps of one
hour each into the context. Even with 30% of values missing, the error rose by
only about 0.2 percentage points. You pass the gaps as NaN, there is nothing
to fill in.
Without a GPU: Timestral on the CPU
Many companies have no GPU in their servers, and Timestral does not need one. We measured run times on an AMD Ryzen 9 5900X with eight threads, in fp32 and without further optimization, for a 96-step forecast:
| Model | 1 series, 2048 steps | 32 series, 2048 steps | 1 series, 32,768 steps |
|---|---|---|---|
| 5M | 13 ms | 0.16 s | 0.11 s |
| 28M | 37 ms | 0.54 s | 0.25 s |
| 209M | 174 ms | 2.7 s | 1.5 s |
On a server CPU with bfloat16 and torch.compile, Mistral reports 17 ms for
5M and 113 ms for 209M. On an RTX 4080, the 209M model took 0.19 seconds for 64
series.
So for a daily forecast of a few hundred series, an ordinary server is
enough. In our tests the 28M model was often close to the large models and is
five times faster than 209M on the CPU. Covariates increase compute time
considerably, in the Rossmann test on the GPU by a factor of nine. Factor that
into your planning. torch.set_num_threads() sets how many cores PyTorch
uses.
Limitations and pitfalls
License
The code is licensed under Apache 2.0, the weights under a "Modified MIT" license. The change compared to MIT: you may not use the weights if the global consolidated monthly revenue of your company or your employer exceeded USD 20 million in the previous month. This also applies to derived models. Larger companies need a commercial license from Mistral or use Mistral's hosted services. Check this before you start a project.
No fine-tuning in the open code
The repository only contains inference. Mistral offers fine-tuning on your own data through its Forge platform, including on your own hardware. With the open code, you stay with zero-shot plus covariates.
Unknown training data
Which real datasets were used in training has not been published. So test on data created after October 15, 2026, or on your own non-public data.
Tricky covariates
A user on GitHub documented a case in which the upper quantiles explode: q99
in the thousands for items that sell two units a day. The trigger was a price
that is constant throughout the context and only changes in the forecast
horizon. Per-row normalization turns a few cents into a huge jump. The user
sets such rows to NaN. Together with our temperature finding, this means:
test every covariate on its own.
Installation and platform
The package is not on PyPI, you install it from Git. On Apple GPUs (MPS) it currently fails because the statistics are computed in float64. On CPU and NVIDIA GPUs everything ran without problems for us.
No timestamps
Timestral only sees the order of the values. Calendar, opening hours and holidays have to be passed as covariates, as shown in the section on holidays.
When Timestral is worth it
In our tests, Timestral replaced a set of classical methods with one model and a few lines of code. Without any training, it beat MSTL, Chronos-2 and a LightGBM baseline as soon as we gave it the necessary extra information. It pays off most in these cases:
- You have many series and little time per series, such as stores, products, meters or sensors.
- You need a good baseline forecast quickly, against which later specialized models have to prove themselves.
- You need uncertainties, for order quantities, staffing or grid reserves, and not just a point value.
- You want to start without MLOps effort: no training, no retraining, one model for all series.
A well-built trained model can still be better, especially when rare events matter a lot and many years of history are available. Our LightGBM won without the day type trick because it knew holidays from three years of data. The most important work step therefore shifts: instead of training models, you think about which information the model cannot see in the context and pass it in as a covariate.
To get started, we recommend:
- Start with the 28M or 123M model and the plain history, and measure against a naive value and your existing method.
- Add calendar and known events as categorical covariates.
- Test every further covariate on its own and keep what helps.
- Check the coverage of the quantiles on your own data before planning with them.
If you want to build forecasts for energy, sales or machine data and need support, get in touch.
Sources
- Mistral AI: Introducing Timestral, October 15, 2026
- Mistral AI: Timestral on GitHub and Timestral-209M model card
- Technical report: Timestral, arXiv 2610.37585
- License of the weights: Modified MIT License
- GitHub issues on training data, exploding quantiles with price covariates and memory use and MPS
- Electricity and solar data: Energy-Charts by Fraunhofer ISE
- Weather forecasts: Open-Meteo Historical Forecast API
- Rossmann data: fev-bench datasets on Hugging Face, originally from the Kaggle competition Rossmann Store Sales
- Sensor data: Numenta Anomaly Benchmark
- Chronos-2: amazon/chronos-2
Interested in building high-quality AI agent systems?
We prepared a comprehensive guide based on cutting-edge research for how to build robust, reliable AI agent systems that actually work in production. This guide covers:
- Understanding the 14 systematic failure modes in multi-agent systems
- Evidence-based best practices for agent design
- Structured communication protocols and verification mechanisms



