Mistral releases Timestral, a 209M model for forecasts

Important

Lab claim

On October 15, 2026, Mistral released Timestral in four sizes, from 5 million to 209 million parameters. The license on the post is a modified MIT.

How to read this claim

Measurement — The technical report is not on the page. No independent rerun is cited.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — The technical report is not on the page. No independent rerun is cited.

Keep the habit · field card with prices and worked checks

What changed

Before this post, a new Mistral release in the briefing was a language model. Timestral is a forecasting model, with a largest size of 209 million parameters and a CPU timing from Mistral.

What it introduces

A four-size forecasting model that Mistral says runs on CPU, with past and future covariates and 21 forecast quantiles.

Who benefits Anyone who forecasts sensor streams, demand, or other numeric series and can compare a new model with the forecast they already trust.

Try this

If you already score a forecast, run the size that fits your CPU budget and compare it with that score. Treat the leaderboard ranks as Mistral's until the technical report is out.

What we do not know

The technical report is not on the page. No independent rerun is cited.

Evidence

Primary Mistral. Reported by Mistral. The leaderboard ranks and the CPU timings are Mistral's. The technical report is not published yet.

Independent The ranks on GIFT-Eval, TIME, and fev-bench are Mistral's. Mistral says the results were submitted to two of those leaderboards. The technical report is not published on this page.

Sources

Mistral Primary · October 15, 2026

The record

What happened

On October 15, 2026, Mistral released Timestral, a family of time-series models at 5 million, 28 million, 123 million, and 209 million parameters. The post says the largest size ranks first on GIFT-Eval, TIME, and fev-bench among pretrained zero-shot models, and that the results were submitted to the GIFT-Eval and TIME leaderboards. On a server CPU with 8 threads, Mistral measures 10 milliseconds for the 5 million parameter model and 140 milliseconds for the 209 million parameter model, forecasting one day ahead from three weeks of 15-minute history. The license on the post is a modified MIT. Context is 32,768 time steps.

Why it matters

Most of this week's models answer in language. Timestral forecasts a numeric series, and Mistral says the 209 million parameter model does not need a GPU for the timing above. That is a different job from Beam or a frontier chat model. The first-place claim is Mistral's, from one checkpoint, and the technical report is not on the page yet.

Mistral says the 28 million parameter model beats Chronos-2, which it describes as about four times larger, on GIFT-Eval and TIME. That comparison is Mistral's.

The post says a GitHub repository is available so the model can run on your own data. It does not print the repository address on the page we read.