Mistral Timestral tested: forecasting AI without a GPU

TL;DR. On October 15, Mistral released Timestral: four AI models that extend series of numbers, such as sales, electricity demand or machine data. They run on an ordinary processor without a graphics card, and without any prior training on your data. We had them forecast German electricity demand for 14 days. The second-smallest model was off by 2.15% on average, the rule of thumb "same as last week" by 3.39%. One day-ahead forecast took 32 milliseconds. Good forecasts no longer need a data centre. What matters now is clean data and a look at the licence, because above 20 million dollars in monthly revenue it is no longer free.

What Mistral has released

Timestral is Mistral's first model family for time series, meaning numbers that come in at fixed intervals: sales per day, consumption per quarter hour, temperature per minute. It does not write text, it continues the curve. The key facts according to Mistral's announcement and the model cards:

Four sizes. 5, 28, 123 and 209 million parameters. As files, that is 21 to 837 megabytes.
Usable without training. You feed in the past, the model returns the forecast. Up to 32,768 past values, up to 1,024 steps ahead in a single pass.
A range instead of one number. For every point in time it returns 21 levels, from the 1% to the 99% value. That shows how certain the forecast is.
Extra knowledge possible. Public holidays, promotions or weather forecasts can be passed in as series of their own.
Runs in-house. On your own server, in your own cloud or fully offline. The code is freely available on GitHub, the models are on Hugging Face.

On the three most important public benchmarks for forecasting models, GIFT-Eval, TIME and fev-bench, Mistral says the largest model ranks first among models that work without additional training. The lead is small, though: on the MASE error measure in GIFT-Eval it scores 0.658, against 0.661 for the runner-up and 0.667 for Google's TimesFM-3.

Our test: German electricity demand

Benchmarks are one thing. We wanted to know how Timestral handles data it cannot possibly know. So we took electricity demand in Germany as published by the Federal Network Agency on SMARD, one value per quarter hour. For every day from October 3 to 16, each model got the three weeks before, 2,016 values, and had to forecast the next 24 hours. No holidays, no weather, no additional training. The computer was a Mac with an Apple M4 Max using eight processor cores, no graphics card. All test days fall after training, October 15 and 16 even after the release.

The opponent was not another AI, but what many planners already do: take yesterday's value, or the value from the same weekday a week earlier.

Electricity demand in Germany on October 16, gigawatts

actual demandTimestral-28M forecast80% likely rangeRule of thumb: same as last week

Source: Federal Network Agency, SMARD, grid load Germany. Forecast from the three weeks before October 16, our own measurement. On that day Timestral was off by 2.1% on average, the rule of thumb by 2.9%.

Across all 14 days, the gap looks like this:

Mean error per day-ahead forecast, October 3 to 16

Same as yesterdayRule of thumb

7.81%

Same as last weekRule of thumb

3.39%

Timestral-5M21 MB · 12 ms per forecast

2.43%

Timestral-28M111 MB · 32 ms per forecast

2.15%

Timestral-123M490 MB · 114 ms per forecast

2.23%

Timestral-209M837 MB · 147 ms per forecast

2.26%

Rule of thumbTimestralBest result

Our own measurement, 14 days of 96 quarter hours each. Error = mean deviation from the actual value in percent. Times on a Mac with an Apple M4 Max, eight processor cores, no graphics card.

Four things stand out:

Small is enough. The 28-million model was the most accurate, the larger ones came close behind. With a series this regular, more size buys nothing.
The range holds. Between the 10% and the 90% value, 80% of the actual values should fall. For the 28-million model it was 90%, for the others 83%. The band errs on the cautious side rather than being too tight.
It does not know holidays by itself. On October 3, a public holiday in Germany, the rule of thumb was slightly better at 3.2% than the model at 3.4%. Timestral only knows about special days what you pass in.
It is fast. 12 milliseconds for the smallest model, 147 for the largest. That matches Mistral's figure of 10 to 140 milliseconds.

A test with one series over 14 days is not proof. The electricity demand of a whole country is well-behaved, it follows the day, the week and the weather. Sales of a single spare part jump around far more.

The small print

The licence has a revenue cap. According to the licence, you may not use the models if your company made more than 20 million dollars in worldwide revenue in the previous month, counted for the whole group. That is around 240 million dollars a year. Above that, the options are a licence from Mistral or Mistral's own service.
The numbers come from Mistral. Mistral has only announced its technical report so far. Researchers warn that public test data can end up in the training of such models (study). Mistral states that no test series was seen in training and that it trained mostly on synthetic series.
The lead is narrow. TimesFM-3 from Google and Chronos-2 from Amazon are close behind. If you already use one of them, there is no need to switch.
It is not a planning system. Timestral delivers numbers, not purchase orders. Pulling data from the ERP, filling gaps, feeding results back and monitoring them is still work. Mistral itself writes that real data shifts over time and needs readjusting.

What this means for a mid-sized company

Until now, forecasts were either manual work in Excel or a project: someone who builds and maintains a separate model for every series, or expensive planning software. Models like Timestral turn this around. A file of just over 100 megabytes, no training, an existing server, and the data never leaves the building. That lets a forecast run alongside every series: sales per item, spare parts demand, electricity and gas purchasing, shift planning, sensor values of a plant.

The bottleneck moves from the model to the data. Gap-free time series, clean item master data and known special days now decide the quality. If you have those, you can test in a few days what used to take a project.

What you should do now

Pick a series you already plan today. For example sales of your 50 most important items, or the electricity demand of one plant. To start with, the more regular, the better.
Test against your own rule of thumb. Backwards, with weeks the model cannot have seen. The yardstick is not the perfect forecast, but what your planners do today.
Use the range. Do not plan by the middle value. Decide how much risk you are willing to carry. The 90% value is a good starting point for safety stock.
Pass in special days. Public holidays, promotions, company shutdowns and price changes belong in as series of their own. Without them, the model guesses on exactly the days that cost the most.
Check the licence. If your group's revenue is above 20 million dollars a month, settle the licence with Mistral before the first real use.

Timestral does not make forecasts error-free. But it makes them cheap enough to try wherever people still estimate today. If you want to know which of your series are suited to it, we are happy to talk.