TL;DR: Mistral released Timestral, its first family of time-series forecasting models, in four sizes from 5M to 209M parameters under a modified MIT license. They forecast data they’ve never seen without training first, give a range of likely values rather than one number, and don’t need a GPU: Mistral says the largest model takes 140 ms on an 8-thread server CPU for a one-day forecast. Code is on GitHub.

What happened

On October 15, Mistral announced Timestral, “our first family of time-series foundation models.” A time-series model forecasts numbers that change over time, such as sales, sensor readings or server load.

According to Mistral’s announcement:

  • Four sizes. 5M, 28M, 123M and 209M parameters, all with the same design. Bigger is more accurate; smaller is faster.
  • No GPU needed. “On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.”
  • Ranges, not one number. It predicts 21 quantiles, so you see how uncertain a forecast is.
  • Outside factors. It can use related series and “known-future covariates” such as production schedules, planned shutdowns or weather forecasts, plus categories like day of the week.
  • Benchmarks. Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench among pre-trained, zero-shot models. These are Mistral’s own results; Mistral says they’ve been submitted to the GIFT-Eval and TIME leaderboards.
  • License and deployment. Modified MIT license; on-prem, VPC or air-gapped. Fine-tuning on your own data goes through Mistral Forge.

Why it matters

Forecasting used to mean fitting a separate model to each series or paying for a platform. A small, openly licensed model that works on new data straight away and runs on an ordinary CPU makes it practical for a small team to forecast demand, staffing or stock from a spreadsheet export, on its own server, without GPU costs.

Keep the caveats in mind. Mistral’s examples are industrial and financial, its benchmark claims come from Mistral, and the full technical report hasn’t been published yet. Check the license file before commercial use, since it’s a modified MIT license, not plain MIT.

What to do

  1. Try the smallest model first. Mistral has published a GitHub repository for running Timestral on your own data; the 5M or 28M model is enough to test the workflow.
  2. Backtest before trusting it. Hide your last few weeks of data, forecast them, and compare with what really happened.
  3. Add the things you already know. Holidays, promotions and opening hours are the kind of known-future inputs Timestral is built to use.
  4. Use the range. Plan stock or staff on the upper quantile, not just the middle forecast.
  5. Read the license. Check the modified MIT terms in the repository before you build it into a product.

Sources & method

Sources

Method. Last checked October 18, 2026. Details and quotes come from Mistral’s announcement and model pages. Benchmark and speed figures are Mistral’s own. We haven’t tested Timestral. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.