Mistral Opens Timestral, 209M Time-Series Model Beating TimesFM-3

Mistral's largest time-series model has 209M parameters. According to Mistral's launch post, that model "ranks first on all three of the field's main public benchmarks, GIFT-Eval, TIME, and fev-bench, outperforming the previous state of the art with only two-thirds of the model parameters." In the family-sizing section, Mistral is more specific: "Our 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters."

Timestral ships in four sizes — 5M, 28M, 123M and 209M — under a Modified MIT licence, with on-prem, VPC and air-gapped deployment. The models take a context of 32,768 time steps, accept multivariate targets with past-only and known-future covariates, and output probabilistic forecasts over 21 quantiles. "The objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS," the post says.

The other claim worth pulling out is the hardware floor. "Timestral does not need a 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 model." Mistral frames the pitch around enterprise data types: "Time series is one of the most common data types in enterprise, from sensor telemetry to observability metrics. Forecasts built on it drive decisions in capacity planning, energy, supply chains, risk and maintenance."

Fine-tuning is routed through Mistral Forge, "the same training pipeline we use for our language models." This is Mistral's eleventh story in our Mistral tracker over the past 90 days, and the first that is not a language model. The launch post lists no individual authors and carries no external quotes; the benchmark claims are the company's own.