Timestral is Mistral AI's new family of time-series foundation models, ranging from 5 million to 209 million parameters. The largest variant is the company's flagship for zero-shot forecasting, where it outperforms Google's TimesFM-3 on the GIFT-Eval, TIME, and fev-bench benchmarks.
According to the announcement, the 209M model achieves these results without fine-tuning, and it can run on CPU rather than requiring specialized accelerators. That combination of strong benchmark performance and modest hardware requirements could make the model attractive for production forecasting tasks.
Timestral's release adds another option to a growing field of general-purpose time-series models, which aim to predict future values across domains without task-specific training. Mistral positions Timestral as a direct competitor to Google's TimesFM-3, and the benchmark results suggest it holds an edge in the zero-shot setting.