Mistral releases compact models for time-series forecasts
In brief
Mistral AI has released Timestral, its first family of foundation models for time-series data. According to the company, the largest model has 209 million parameters and ranks first on the GIFT-Eval, TIME and fev-bench benchmarks among comparable pre-trained models. The four sizes, from 5 million to 209 million parameters, produce probabilistic multivariate forecasts, accept known future factors, and can run on a standard CPU.
What it changes for an SME
If your firm already records regular series such as sales, inventory, energy use or machine sensors, Timestral is a new option to test, including on series the model has not been trained on, according to Mistral. Inference is said to run on an ordinary server CPU, so you may avoid graphics hardware. Smaller versions are faster, and fine-tuning is offered through Mistral Forge. Operational use still requires a check on your own data, because public benchmark ranks do not guarantee a gain for your series.
What you can do
- List the recurring series you already store and which planning or maintenance decisions they feed.
- Check whether a CPU-only forecast could sit in your current monitoring or planning tools.
- If you trial a public model, compare it with your current method on a held-out period before you rely on it.
Let's talk about the time you could save.
In 20 minutes, we find where AI could save you time or money. Free, no commitment.