Timestral: Mistral release time-series foundation models

What happened

Mistral launched Timestral, a set of four forecasting models. Its largest model has 209 million parameters, which are the numbers inside an AI model that it learns during training. The company says this model ranks first on three main public tests for forecasting. It beats the previous best model while using only about two-thirds of its size.

Key numbers

• Four sizes: 5M, 28M, 123M and 209M parameters.
• The 209M model ranks first on GIFT-Eval, TIME and fev-bench.
• It uses about two-thirds of the parameters of the previous best model.
• On a standard server CPU, a one-day forecast from three weeks of 15-minute data takes 10 ms with the 5M model and 140 ms with the 209M model.
• The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME.

Why forecasting models are changing

Older forecasting tools, such as ARIMA, learn from one data series at a time. Deep learning models can learn across many series, but they are usually trained from scratch for each dataset. Foundation models are trained once on a large, varied collection of data. They can then make forecasts on new data they have never seen, which helps when data is scarce.

Why it matters

Time series data is common in business. Factories, energy firms, banks and supply chains use forecasts to plan capacity, spot broken equipment and manage risk. A smaller model that runs well on ordinary CPUs, without a GPU, can be cheaper and easier to use. Mistral also says the models support covariates, which are extra factors such as holidays or prices that can help a forecast.

What to watch

• Whether the leaderboard results hold up as other teams test Timestral.
• How companies use fine-tuning through Mistral Forge to adapt the models to their own data.
• Whether rival forecasting models quickly match these results.