Mistral released Timestral on Thursday, October 15, 2026 — a family of downloadable forecasting models that run on a plain CPU — and it matters because it brings foundation-model forecasting to any team with a metrics table, provided their company stays under a $20 million-a-month revenue cap written into the license.
October 15, 2026 — The French lab’s newest release does not generate text or images at all. It reads a column of numbers and predicts where it goes next.
What happened
Mistral shipped Timestral, a family of four time-series foundation models sharing one architecture: Timestral-5M, Timestral-28M, Timestral-123M, and Timestral-209M, named for their parameter counts. They are zero-shot forecasters — you hand them historical data and they produce a forecast without any training on your data first. The weights are downloadable from Hugging Face, the inference code is on GitHub, and Mistral’s documentation lists Timestral 1.0 as a public preview with no API pricing, so for now you run it yourself.
The details that matter
The largest model reads up to 32,768 past steps and forecasts up to 1,024 steps ahead in one pass. Each forecast step comes back as 21 quantiles, from the 1st to the 99th percentile, so every prediction carries a range of likely values instead of a single number. It accepts one series or several at once, plus known events such as holidays or planned promotions as covariates.
Speed is the other headline. Mistral says the models do not need a GPU at inference: on a server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute data reportedly takes 10 milliseconds for the 5M model and 140 milliseconds for the 209M. Mistral also says the 209M ranks first on the field’s main public forecasting benchmarks — on its model card it posts a GIFT-Eval MASE score of 0.658, ahead of Chronos-2’s 0.698, with lower being better — but those are the company’s own numbers, the margins over the runner-up are small, and the technical report is still listed as coming soon.
The license catch
The weights ship under what Mistral calls a Modified MIT license. The modification is a revenue ceiling: a company whose global monthly revenue exceeded $20 million in the preceding month is not authorized to use the weights — including derivatives and fine-tuned copies — without a separate commercial license from Mistral. The cap applies to every checkpoint, so picking a smaller model is not a way around it. The inference code itself is Apache 2.0; only the weights carry the restriction.
Why it matters
Forecasting is unglamorous, essential work: capacity planning, inventory, energy demand, anomaly alerts. A credible zero-shot model that runs in milliseconds on a CPU removes both the GPU bill and the per-team training project. But the license draws a bright line through the market — startups and mid-size companies can build on Timestral freely, while the large enterprises most likely to want on-premise, air-gapped forecasting are exactly the ones who must negotiate a commercial deal first. Check which side of the line your employer sits on before anyone builds on it.
Frequently asked questions
What is Mistral Timestral?
A family of four open-weight time-series foundation models (5M to 209M parameters) released October 15, 2026. They forecast future values from historical numeric data with no task-specific training.
Does Timestral need a GPU?
No. Mistral reports CPU inference in the 10–140 millisecond range for a day-ahead forecast on an 8-thread server CPU, depending on model size.
Who cannot use Timestral for free?
Companies with more than $20 million in global monthly revenue. The Modified MIT license withholds authorization above that line; larger companies need a commercial license or Mistral’s hosted services.
Sources: Mistral AI, Hugging Face model cards, GitHub, DEV Community