MistralAI Launches Timestral-28M for Zero-Shot Time Series Forecasting

MistralAI has released the Timestral-28M, a time series foundation model designed for zero-shot forecasting, featuring 28 million parameters. This patch-based transformer model incorporates time and variate attention across 12 layers with an embedding dimension of 384 and supports a context length of up to 32,768 time steps. The model is capable of forecasting up to 1,024 time steps in a single forward pass and outputs 21 quantiles. It can be run locally using PyTorch on both CPU and GPU. According to the model card, Timestral-28M offers a balance between model size and forecasting performance, achieving notable scores on benchmarks such as GIFT-Eval and TIME. The model is available under a Modified MIT license, with the restriction that it must not be used in a manner that violates third-party rights.