Mistral AI unveils Timestral: a new forecasting model family outperforming google's timesfm-3

Mistral AI has launched Timestral, a family of time series forecasting models that outperforms Google’s TimesFM-3. The model family includes four zero-shot forecasting models, ranging from 5M to 209M parameters. The largest model, Timestral-209M, ranks first in GIFT-Eval, TIME, and fev-bench without using a GPU. It provides probabilistic forecasts on standard CPUs with open weights and inference code. Timestral is a causal, patch-based transformer with time and variate attention. It transfers patterns across domains, forecasting multivariate targets jointly. The 209M model uses 24 layers and an embedding dimension of 768, with each patch covering 32 time steps. It forecasts up to 1,024 steps in a single forward pass. Mistral highlights its efficiency, with the 5M model completing a 1-day forecast in 10 ms and the 209M model taking 140 ms. The 28M model is five times faster than the 209M model. The model is available via pip, requiring Python 3.10+ and PyTorch 2.4+. Weights are hosted on Hugging Face, with sizes ranging from 21 MB to 490 MB. Mistral’s documentation lists Timestral 1.0 209M as Public Preview, with fine-tuning options through Mistral Forge. The model is deployable on-prem, in a VPC, or in an air-gapped environment.