Mistral launches Timestral, small forecasting models that run on a CPU Other

Mistral has released Timestral, its first family of time-series foundation models. They make probabilistic, multivariate forecasts and run on a CPU with no GPU. Mistral announced the models on October 15, 2026.

Four sizes, one architecture

All four models share the same architecture and training recipe.

Latency is for a one-day forecast from three weeks of 15-minute data on a standard server CPU with 8 threads.

Benchmarks

Mistral's own numbers, for pretrained zero-shot models:

• The 209M model ranks first on GIFT-Eval, TIME and fev-bench, on all five metrics Mistral reports.
• It beats TimesFM-3 and Falcon-2.0 with fewer parameters.
• The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME.
• Results come from a single checkpoint, without ensembling. Mistral has submitted them to the GIFT-Eval and TIME leaderboards.

A technical report with full results is promised but not out yet.

What it can do

• Context of up to 32,768 time steps.
• Output: a probability forecast across 21 quantiles.
• Inputs: several target series, past-only covariates and known-future covariates, including categorical ones like day of week.
• Most training data is synthetic, followed by post-training on a mix of real and synthetic data.

License and access

• Model weights: Modified MIT license, on Hugging Face.
• Code: Apache 2.0, on GitHub.
• Mistral Forge offers fine-tuning on private data, custom covariates and deployment on-prem, in a VPC or air-gapped.

Sources

• Mistral: Timestral announcement
• Timestral on GitHub
• Timestral-209M on Hugging Face
• Mistral docs: Timestral 1.0