Timestral launches compact forecasting family

Timestral makes a strong case that forecasting quality depends more on architecture, training data, and deployment fit than raw parameter count.

• –Mistral says Timestral-28M beats the roughly four-times-larger Chronos-2 on GIFT-Eval and TIME, while Timestral-209M surpasses TimesFM-3 and Falcon-2.0 with fewer parameters.
• –The four-size lineup creates a practical accuracy-latency tradeoff: Mistral reports 10 ms CPU inference for 5M and 140 ms for 209M on a representative workload.
• –Multivariate modeling, future covariates, and 21 quantile outputs make the models useful for capacity planning, supply chains, maintenance, and risk—not just point predictions.
• –The code is Apache 2.0 and the weights use a Modified MIT license, enabling local, VPC, and air-gapped deployment. [GitHub repository](https://github.com/mistralai/timestral)
• –Benchmark leadership is promising but remains primarily a vendor-reported claim; production teams should validate performance on their own domains and baselines.