# Timestral: Mistral's time-series foundation models | Mistral

The models range from 5M to 209M parameters, run inference on CPUs, and top the GIFT-Eval, TIME, and fev-bench zero-shot leaderboards.

- Released in four sizes: 5M, 28M, 123M, and 209M parameters, supporting probabilistic and multivariate forecasting.
- Supports both past-only and known-future covariates using covariate-aware attention that prevents future target leakage.
- Runs on standard 8-thread server CPUs in 10 ms for the 5M model and 140 ms for the 209M model for a one-day forecast.
- Pre-trained largely on synthetic multivariate data before post-training with real data using pinball loss across 21 quantiles.
- Available as code on GitHub and accessible for custom fine-tuning and domain adaptation through Mistral Forge.

## Why it matters

Organizations forecasting telemetry, supply chains, or financial metrics can deploy zero-shot time-series models directly on CPUs without training models from scratch.

## Sources

- [Mistral: Timestral: Mistral's time-series foundation models | Mistral](https://mistral.ai/news/timestral)

---

Summarized by dstilled on 2026-10-15. https://dstilled.ai/story/a040149b-ea61-4d20-a277-8c0cb799389d