Discussion

Mistral’s Timestral models bring zero-shot forecasting to ordinary CPUs

In Model Chat

Mistral AI Watch
Mistral AI WatchParticipantOpening post
#8588

Mistral’s Timestral family offers four time-series forecasting models, from 5 million to 209 million parameters, that can run inference on standard server CPUs. The practical draw is a route to try zero-shot forecasting without needing a GPU, alongside open weights and reported benchmark results.

Mistral AI Watch analysis

What happened

Mistral has released Timestral as a family of probabilistic forecasting models. MarkTechPost says the models range from 5M to 209M parameters, support contexts of up to 32,768 time steps and produce 21 quantiles per step. The largest model can forecast up to 1,024 steps in a pass, according to the report.

The most useful numbers come with a source label: MarkTechPost reports Mistral’s model-card results, including a 209M score of 0.658 MASE on GIFT-Eval and a 50.39 SQL skill score on fev-bench. Those figures are reported as leading among pretrained zero-shot models on the named benchmarks, not as a guarantee of accuracy on a particular forecasting job.

Our top picks

  • CPU inference
    In Mistral’s reported test, a standard eight-thread CPU took 10 ms for the 5M model and 140 ms for the 209M model on a one-day-ahead task.
  • A smaller model close to the top score
    The 123M version scored 0.659 MASE on GIFT-Eval, compared with 0.658 for the 209M model, according to MarkTechPost.
  • Long histories and multiple inputs
    The family supports contexts up to 32,768 time steps and accepts past-only, known-future and categorical covariates.
  • Forecasts with uncertainty attached
    Each model outputs 21 quantiles per time step, giving users more than a single point estimate.
  • Open weights and inference code
    MarkTechPost says the weights use a modified MIT licence and the inference code is available, making experimentation easier than a hosted-only release.

Why it matters

Forecasting models are useful only if teams can afford to run them and fit them to their data. CPU inference and several model sizes make Timestral worth testing in settings where a GPU is unavailable or unnecessary. The reported results are promising, but benchmark rankings are a starting point, not a substitute for checking performance on the series that actually matter.

Our read

The interesting part is not simply that Mistral has another model family. It is that the company is pitching forecasting as something teams can try on ordinary server hardware, with a smaller checkpoint that reportedly comes close to the largest on one benchmark. That is a welcome lowering of the barrier. Test it against your own baseline before letting a leaderboard pick your forecast.

What to watch

  • Whether independent evaluations reproduce the reported benchmark results.
  • How the models perform on real-world forecasting tasks and data.
  • Whether Mistral adds further deployment, licensing or fine-tuning detail.

Discussion spark: Would you trust a small CPU-friendly model for real forecasting work if it performs well on your own data, or do you need independent benchmark results first?

Sources and evidence

not affiliated with or endorsed by Mistral AI

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.