The AI Toolcast.

AI tool news. Practical perspective.

Follow with RSS

Mistral releases Timestral forecasting models for CPUs and GPUs

Mistral's new models forecast numerical data locally. Developers can download the weights, but the inspected licenses for the smallest and largest models include a revenue restriction.

BriefModels2 min read

By AI Toolcast deskSource review · AI-assisted
View sources
In this article
  1. Quick answers
  2. Local forecasts with uncertainty estimates
  3. Code and weights have different terms
  4. Why data teams should care
  5. Sources & method

Quick answers

What does Timestral forecast?

Numerical time series, including related measurements, with support for inputs known in advance, such as production schedules.

Source 1 Source 3

Is a GPU required?

No. The official inference package supports CPUs and GPUs; its installation instructions require Python 3.10 or later and PyTorch 2.4 or later.

Source 2

Are the downloaded weights unrestricted?

No. The inspected 5M and 209M licenses impose a $20 million monthly company-revenue threshold and require preserving attribution and permission notices.

Source 4 Source 5

The practical takeaway

Timestral gives data teams a local forecasting option; check the weight license separately from the code license.

Local forecasts with uncertainty estimates

Mistral released Timestral on October 15, offering four forecasting models from 5 million to 209 million parameters. Developers can download weights from Hugging Face and run them through a public Python package on CPUs or GPUs.

The models process numerical histories and produce 21 quantiles: estimates across a distribution of possible outcomes. They can combine related measurements and incorporate inputs known ahead of time, such as production schedules.

Source 1Source 2Source 3

Code and weights have different terms

The inference code uses Apache 2.0. The weights have separate Modified MIT terms. The 5M and 209M licenses we inspected do not grant permission to users whose company or employer exceeded $20 million in global consolidated revenue in the preceding month. Those users can request a commercial license, which Mistral may grant at its discretion.

The 209M Hugging Face page currently lists no Inference Provider deployment. For local use, developers install the package and load the weights.

Source 2Source 3Source 4Source 5

Why data teams should care

Our reading: CPU support offers a practical route to evaluate forecasts on existing infrastructure. For capacity planning, a range of possible outcomes can reveal shortfall risks that a single prediction leaves hidden. That makes uncertainty estimates useful, provided they hold up on the team's own data.

Mistral reports leading zero-shot benchmark results, meaning forecasting without task-specific training. We have not reproduced those results; the launch post says a full technical report will follow.

Source 1Source 3

Sources & method

AI-assisted reporting from Mistral's announcement, official repository, model card, and the 5M and 209M weight licenses, with original analysis. We have not run the models or independently verified performance claims.

  1. Mistral: Introducing TimestralPublished October 15, 2026
  2. Mistral: Timestral inference repository and READMEAccessed October 18, 2026
  3. Mistral: Timestral-209M model cardAccessed October 18, 2026
  4. Mistral: Timestral-5M Modified MIT licenseAccessed October 18, 2026
  5. Mistral: Timestral-209M Modified MIT licenseAccessed October 18, 2026

Source pages reviewed October 18, 2026. Analysis and suggested evaluation methods are AI Toolcast’s own.

This article contains no sponsored placement or affiliate links.

Keep reading