Introducing Timestral
mistral ·
Mistral has released Timestral, its first family of time-series foundation models designed for probabilistic and multivariate forecasting. The largest model contains only 209 million parameters yet achieves state-of-the-art results on major benchmarks like GIFT-Eval, TIME, and fev-bench, outperforming significantly larger competitors. This efficiency allows the models to run effectively on CPUs without requiring GPUs during inference. The family includes four sizes ranging from 5M to 209M parameters, balancing accuracy with computational speed for various enterprise needs. Key technical innovations include decoupled horizon forecasting, covariate-aware attention, and a synthetic-first pre-training strategy. These features enable zero-shot forecasting capabilities and support both past-only and known-future covariates for improved prediction quality. Organizations can further adapt the models through Mistral Forge for fine-tuning on proprietary data or deploying in air-gapped environments.
mistral