Mistral releases Timestral, a time-series model family that beats TimesFM-3 at 209M parameters

What was announced?

Mistral AI released Timestral, its first family of time-series foundation models. The family consists of four dense models at 5M, 28M, 123M and 209M parameters. Weights and inference code are open, and the largest checkpoint is 837 MB in fp32.

Why it matters

Time-series models usually require per-dataset training or heavy hardware. Timestral instead forecasts zero-shot on unseen data and needs no GPU at inference. Mistral frames this as probabilistic forecasting on standard server CPUs.

Technical design

It is a patch-based transformer with time attention and variate attention. Each patch covers 32 time steps, and context reaches 32,768 steps. The model outputs 21 quantiles per step and forecasts multivariate targets jointly. Mistral lists four contributions: decoupled horizon forecasting, covariate-aware attention, synthetic-first data and two-stage training.

Benchmark results

Per Mistral's model card, the 209M version scores 0.658 MASE and 0.446 CRPS on GIFT-Eval, and 0.626 MASE and 0.522 CRPS on TIME. Its fev-bench SQL skill score of 50.39 edges out TimesFM-3 at 48.99. The GIFT-Eval gap is narrow, however: 0.658 versus 0.661.

Deployment

Mistral says the models can be deployed on-prem, in a VPC or air-gapped via Mistral Forge. The same PyTorch code runs on CPU or CUDA.