Mistral's Timestral time-series models: strong point forecasts, but poorly calibrated intervals in one test

Mistral released Timestral, a family of four time-series foundation models ranging from 5M to 209M parameters, with probabilistic forecasts and a claimed top zero-shot ranking on three forecasting benchmarks. A developer who downloaded the weights confirmed the parameter counts and reproduced the claimed CPU latency on a 2-vCPU machine, finding roughly 52 ms for the 5M model against Mistral's 10 ms figure on an 8-thread server. On a classic airline passenger dataset, the 5M model's median forecast beat seasonal naive and drift baselines on error. However, its 80% prediction intervals covered only 4 of 24 held-out months, and all medians fell below actual values, indicating poor uncertainty calibration.