Mistral presents Timestral, a forecasting model with 209M parameters it ranks first on three benchmarks

Mistral's first time-series models run on a server CPU, from 10 ms at 5M parameters to 140 ms at 209M. The release names no CPU, and the full technical report is still to come.

AI & Compute 2 minute read

Timestral is Mistral’s first family of time-series foundation models, released on October 15 in four sizes of 5M, 28M, 123M and 209M parameters. In its announcement, Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench among pre-trained, zero-shot models, ahead of the previous state of the art with only two-thirds of its parameters. The results come from a single checkpoint without ensembling and have been submitted to the GIFT-Eval and TIME leaderboards.

Forecasting a day ahead takes 10 ms with the 5M model and 140 ms with the 209M on a server CPU

Mistral states that Timestral needs no GPU at inference. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M. The 28M model is five times faster than the 209M one, and the 123M model comes within 0.2% of the 209M model’s score at three quarters of its latency. Mistral says its 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME.

Every size takes covariates, returns 21 quantiles and reads a context of 32,768 time steps

Every model accepts multivariate targets together with past-only and known-future covariates, including categorical ones such as a machine’s running or idle state, and they return a full set of 21 quantiles instead of a single predicted value. Each model reads a context of up to 32,768 time steps. The quantiles let a team set alert thresholds around the spread of the forecast. The licence is a modified MIT, and deployment can be on-premises, in a VPC or air-gapped. A GitHub repository is published for running Timestral on your own data, and fine-tuning is offered through Mistral Forge, the training pipeline.

No CPU model is given for the latency figures, and the full results are still to come

Latency is quoted without a CPU model, and no GPU timing is given for comparison. A technical report with the full results is still to come, so the training method it only outlines, mostly synthetic data and a covariate-aware attention design, is not yet documented in detail.