Mistral releases Timestral, a model that forecasts future values from past numbers (mistral.ai)

Writing language: KoreanRead in the original language

- Timestral is a family of models that looks at numbers recorded in time order, such as equipment sensor readings or daily sales, and predicts future values. It comes in four sizes, from 5M to 209M parameters, and instead of a single predicted value it shows a range of uncertainty using several quantiles. - Mistral says its largest 209M model finished 1 place on three public benchmarks: GIFT-Eval, TIME, and fev-bench. This is the developer's claim. The comparison was made among zero-shot models that are used without additional training, using a single checkpoint for each. - According to the source, on a 8-thread server CPU, forecasting one day ahead from 3 series recorded at 15-minute intervals took 10ms for the 5M model and 140ms for the 209M model. These speed figures are also conditions measured and reported by Mistral, and the source explains that inference can run without a GPU.

Found on

Hacker News ↗ / 216 votes / 43 comments (checked 2026-10-19 23:30 KST)

Have you run into something like this? Tell us in a line or two what you did. You do not need to know the product.

1 comment

Editorial opinionstartupxo 2d ago

O: A factory operations manager who checks equipment sensor values every day and decides next week's parts inventory could pick one internal record from a public GitHub repository, forecast a past period, and compare the forecast with the actual values. X: 209M model's rank 1 and 140ms inference time come from benchmarks and measurement conditions announced by Mistral, so you still need to check directly whether the same performance holds on your own data.

Writing language: Korean

Related reads

1 comment