Mistral’s Timestral forecasts time series it has never seen, and runs on a CPU

1 hour ago

Mistral released Timestral, its first time-series foundation models: four open-weight sizes from 5M to 209M parameters that forecast data they have never seen, with uncertainty bands, on a CPU. Also: Mistral Large 4 in public preview in the API, with weights due by the end of October. https://mistral.ai/news/timestral/ https://huggingface.co/mistralai/Timestral-209M https://docs.mistral.ai/models/timestral-1-0 https://mistral.ai/news/mistral-large-4/

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Timestral
  2. 0:16A range, for related series
  3. 0:30Known-future inputs
  4. 0:43Runs on a CPU
  5. 0:57Benchmarks
  6. 1:13Mistral Large 4
  7. 1:27Run Timestral
Show transcript

Timestral

Mistral AI

Timestral forecasts time series it has never seen

Mistral’s first time-series models · four sizes · open weights

Timestral hourglass artwork
mistral.ai/news/timestral · 15 Oct 2026

Mistral has released Timestral, its first family of time-series foundation models. Give it sensor readings, server metrics or account flows, and it forecasts what comes next without training on your data first.

A range, for related series

Mistral AI

A forecast with its uncertainty built in

21 quantiles per step · related series forecast together

Timestral model overview: target variates and covariates feed a transformer that outputs quantile forecasts
mistral.ai/news/timestral

Timestral predicts 21 quantiles at every step, so you see how wide the uncertainty is and can set alert thresholds around it. It also forecasts related series together, like a pump’s vibration, temperature and pressure.

Known-future inputs

Mistral AI

Tell it what you already know about tomorrow

Production schedules Planned shutdowns Weather forecasts Machine state: running, idle, maintenance
mistral.ai/news/timestral

You can feed Timestral inputs known ahead of time, such as production schedules, planned shutdowns and weather forecasts. A scheduled load increase then shows up in the forecast instead of tripping an alert.

Runs on a CPU

Mistral AI

Timestral runs on a CPU

One day ahead from three weeks of 15-minute history · 8-thread server CPU

Timestral 5M5M parameters
10 ms
Timestral 209M209M parameters
140 ms
mistral.ai/news/timestral

Timestral doesn’t need a GPU. On an eight-thread server CPU, the smallest model forecasts a day ahead in about 10 milliseconds, and the largest, at 209 million parameters, takes about 140.

Benchmarks

Mistral AI

First on GIFT-Eval, TIME and fev-bench

Among pretrained zero-shot models · one checkpoint, no ensembling

GIFT-Eval GMean MASE: Timestral 0.657, lowest of the top ten
mistral.ai/news/timestral · GIFT-Eval GMean MASE, lower is better

On the three main public forecasting benchmarks, the largest Timestral comes closer to the real values than any other pretrained zero-shot model. The 28 million parameter version beats Chronos 2, about four times its size.

Mistral Large 4

Mistral AI

Mistral Large 4 is in the API now, with open weights due this month

1T parameters52B active · multimodal
Public previewMistral Studio API
$0.68 / $2.09Sale price per M tokens in / out · list $1.36 / $4.18
WeightsBy the end of October
mistral.ai/news/mistral-large-4 · docs.mistral.ai/models/mistral-large-4-0

Mistral Large 4 also arrived this month, with one trillion parameters and 52 billion active. Its preview API is open in Mistral Studio at half price for now, and the weights are planned for the end of October.

Run Timestral

Mistral AI
Timestral icon

All four Timestral sizes are on Hugging Face

Modified MIT licence Public preview Technical report to follow

Setup and usage: the Timestral model card on Hugging Face

huggingface.co/collections/mistralai/timestral · docs.mistral.ai/models/timestral-1-0

All four sizes are on Hugging Face under a modified MIT licence. Timestral is in public preview, with the full technical report still to come, and the model card covers setup.