A 209M model claims three firsts; the report is not out
The visible page date is October 15. The post says Timestral, Mistral's first family of time-series foundation models, is out today. The largest model has 209M parameters and ranks first on the three main public benchmarks, GIFT-Eval, TIME, and fev-bench, with about two-thirds the parameters of the previous best. Among pretrained zero-shot models, it says it ranks first on all five metrics, from a single checkpoint, without ensembling. The results have been submitted to the GIFT-Eval and TIME leaderboards. A technical report with the full results will follow.
The firsts are Mistral's evaluation plus a submission. They are not yet a rank confirmed on a leaderboard page. Structured data on the page also carries July 15. The visible date and the body both say October 15, and the body says the release is today. This piece follows the announcement on the page.
[1]
Four sizes, and the CPU times are tens to low hundreds of milliseconds
The four sizes are 5M, 28M, 123M, and 209M. They share the architecture and the training recipe. Mistral says every size sits on the Pareto front of accuracy versus model size: at a given size, no public model forecasts better. That claim refers to the chart they drew. The post does not turn the chart into a table of scores.
On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 milliseconds with the 5M model and 140 milliseconds with the 209M model. The 28M model is five times faster than the 209M model. The 123M model comes within 0.2 percent of the 209M score at about three quarters of its latency. They also say the 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME, and the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. Beyond zero-shot, one checkpoint, and no ensemble, the post does not list the hardware for those comparisons. Inference does not need a GPU.
[1]Synthetic data first, then a context four times longer
Most of the training data was synthetic. The model is pretrained on purely synthetic data with a context of 8192, then post-trained on a mix of real and synthetic data with a context four times longer. The objective is a pinball loss over 21 quantiles. The model card lists a context of 32,768 time steps, probabilistic forecasts as 21 quantiles, and inputs that can be multivariate targets, past-only covariates, and known-future covariates. The license is a modified MIT license. Deployment can be on premises, in a VPC, or in an air-gapped environment.
The example they give is a plant with thousands of sensors, forecasting vibration, temperature, and pressure together rather than each series alone. That is a description of a use, not a measured error from a plant.
[1]