Timestral: Mistral's time-series foundation models

Family of four open-weight time-series foundation models (5M/28M/123M/209M) for zero-shot multivariate probabilistic forecasting with 32K-step context, running on CPU.

Why it matters

What you could build with it

Does it hold up?

Code is Apache 2.0 and pip-installable; benchmark-leadership claims are vendor-reported with the technical report still pending, and the weights carry a Modified MIT license that bars use above $20M/month company revenue, so enterprises need a commercial license. Too early to judge real-world accuracy until independent evals land.

Built with Timestral

• Timestral: Mistral's Tiny AI That Predicts Numbers (No GPU)youtube · 2-minute plain-English video explainer of the Timestral release: zero-shot forecasting, quantile bands, covariates, CPU inference.
• Mistral's Timestral Runs on a CPU. Its 'Modified MIT' License Stops at $20M a Montharticle · Oct 16, 2026 hands-on writeup: checkpoint table, quickstart code, benchmark margins vs TimesFM-3/Chronos-2, and a quote of the license file showing the $20M/month revenue cap on the weights.
• Mistral's Timestral Runs on a CPU (license deep-dive: Modified MIT with $20M/mo revenue cap)dev.to · Independent writeup testing Timestral's CPU inference claims and unpacking the Modified MIT revenue-cap license term.
• Timestral: Mistral's Tiny AI That Predicts Numbers (No GPU) - video explaineryoutube · Short video walkthrough of Timestral's zero-shot forecasting, probabilistic output, and where the tiny models could land.
• Mistral Timestral Forecasts Business Data on a CPUaiparabellum · News-desk breakdown of Timestral's benchmark claims (GIFT-Eval, TIME, fev-bench) and its multivariate/covariate feature set.

Learn more

official docs →

First spotted on github: source.

More AI models