Mistral AI announced a new line of time‑series foundation models called Timestral. The biggest model has 209 million parameters and is designed to predict data that changes over time, such as sales numbers or sensor readings. Mistral says the 209 M version outperforms Google’s TimesFM‑3 on three benchmark suites – GIFT‑Eval, TIME and fev‑bench – when used without any task‑specific fine‑tuning (zero‑shot). The model can run on a standard CPU, so users don’t need a graphics card to try it.
What the model family includes
Timestral is offered in several sizes, ranging from a lightweight 5 M version up to the 209 M flagship. The smaller models are meant for devices with limited compute, while the largest one targets research and business use cases that need higher accuracy. All versions share the same architecture, allowing developers to switch sizes without rewriting code.
Why it matters
For analysts and engineers who need to forecast trends, Timestral offers a ready‑to‑use model that can run on ordinary hardware, lowering the cost of entry. Its stronger benchmark scores suggest more accurate predictions than the previous Google offering, though independent testing is still pending.
