RELEASE · MODELS · #2048
Mistral releases Timestral family of time-series foundation models (5M–209M params)
Mistral has released Timestral, a family of four time-series foundation models (5M, 28M, 123M, 209M parameters) that deliver probabilistic, multivariate forecasts, support past and known-future covariates, and are optimized for CPU inference. The 209M checkpoint ranks first among pretrained zero-shot entries on the GIFT-Eval, TIME and fev-bench leaderboards, and the family is trained using synthetic pre-training, covariate-aware attention, and long-context learning, with optional fine-tuning via Mistral Forge.
KEY POINTS
- Mistral has released Timestral, a family of four time-series foundation models (5M, 28M, 123M, 209M parameters) that deliver probabilistic, multivariate forecasts, support past and known-future covariates, and are optimized for CPU inference.
- The 209M checkpoint ranks first among pretrained zero-shot entries on the GIFT-Eval, TIME and fev-bench leaderboards, and the family is trained using synthetic pre-training, covariate-aware attention, and long-context learning, with optional fine-tuning via Mistral Forge.
- Timestral brings pre-trained foundation models to time-series forecasting with strong zero-shot performance and efficient CPU inference, which can reduce costs and speed deployment of forecasting across industries.
WHY IT MATTERS
Timestral brings pre-trained foundation models to time-series forecasting with strong zero-shot performance and efficient CPU inference, which can reduce costs and speed deployment of forecasting across industries.