Mistral releases Timestral: time-series forecasting models up to 209M parameters

Mistral AI has launched Timestral, a family of four zero-shot time-series foundation models ranging from 5M to 209M parameters. The largest checkpoint ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench while running inference on standard server CPUs without a GPU. Weights and inference code are open under a modified MIT licence.

• 209M model scores MASE 0.658 on GIFT-Eval and CRPS 0.446
• fev-bench SQL skill score of 50.39 vs 48.99 for TimesFM-3
• Context up to 32,768 time steps, 21 quantiles per step
• CPU inference: 10 ms for 5M, 140 ms for 209M