Mistral releases Timestral, a family of time-series forecasting models from 5M to 209M parameters

Mistral AI

Timestral is Mistral's first family of time-series foundation models, built to produce probabilistic, multivariate forecasts. The four models share one architecture and training recipe, and Mistral says the 209-million-parameter version ranks first on the GIFT-Eval, TIME and fev-bench leaderboards among pre-trained zero-shot models.

• The 28-million-parameter model beats Chronos-2 on GIFT-Eval and TIME, though Chronos-2 is about four times larger.
• On a standard 8-thread server CPU, a one-day-ahead forecast from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
• Context length is 32,768 time steps, and each forecast returns 21 quantiles.
• The GitHub repository and Hugging Face collection are available; the listed licence is Modified MIT, and fine-tuning is offered through Mistral Forge.