Mistral releases Timestral, a family of time-series forecasting models

Mistral released Timestral, its first family of time-series foundation models, on October 15, 2026. Its blog post was shared on Hacker News on October 19, where the thread had 585 points and 111 comments by October 20. The figures below are Mistral's own claims; the technical report has not been published yet.

• Four sizes: 5M, 28M, 123M and 209M parameters. Context length is 32,768 time steps. Output is probabilistic forecasts as 21 quantiles, with multivariate targets and past-only and known-future covariates.
• Mistral says the 209M model ranks first among pre-trained, zero-shot models on the GIFT-Eval, TIME and fev-bench forecasting benchmarks, with about two-thirds of the parameters of the previous best model. It says results have been submitted to the GIFT-Eval and TIME leaderboards.
• Mistral says the models run on a server CPU without a GPU: 10 ms for the 5M model and 140 ms for the 209M model, forecasting one day ahead from three weeks of 15-minute history on 8 threads.
• Mistral says most of the training data was synthetic.
• Licence is listed as "Modified MIT". Mistral says it has published a GitHub repository. Timestral is also offered through Mistral Forge for fine-tuning, including on-prem, VPC or air-gapped deployment.