Mistral AI releases Timestral, a family of time-series forecasting models

AIMistral AI releases Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. The largest model ranks first on GIFT-Eval, TIME and fev-bench among pre-trained zero-shot models, and the 209M model runs in about 140 ms per forecast on a standard server CPU. Timestral is available on-prem, in a VPC or air-gapped, with fine-tuning through Mistral Forge, under a modified MIT license.

AIMistral AI releases Timestral-28M, a 28M-parameter time series foundation model for zero-shot forecasting on Hugging Face. The model forecasts up to 1,024 time steps in one pass from a 32,768-step context, and posts a GIFT-Eval MASE of 0.678 and a TIME MASE of 0.638. It is part of a family that also includes 5M, 123M, and 209M versions, released under a Modified MIT license.

AIMistral releases Timestral-209M, a 209M-parameter time series foundation model for zero-shot forecasting, with a family that also includes 5M, 28M, and 123M versions. The model accepts up to 32,768 historical time steps and forecasts up to 1,024 steps per forward pass, outputting 21 quantiles. On the GIFT-Eval, TIME, and fev-bench benchmarks, it posts MASE scores of 0.658 and 0.626 and an SQL skill score of 50.39 on fev-bench, which Mistral describes as state-of-the-art zero-shot accuracy.