Mistral released Timestral on October 15, its first family of time series foundation models. Four checkpoints span 5M to 209M parameters, and the largest ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench.
All of it runs on ordinary server CPU. The 209M weights occupy 837 MB in fp32, and a GPU is optional rather than required, which puts the family well below the hardware floor for most foundation models.
Why small is the point here
Classical tools like ARIMA fit one series at a time. Deep models such as PatchTST usually train from scratch on each dataset. Timestral instead transfers patterns across domains and forecasts data it has never seen.
Each model emits 21 quantiles per step, handles multivariate targets jointly and accepts past-only and known-future covariates. Information flows only from covariates to targets, which blocks future leakage by construction.
The architecture is a causal patch transformer with separate time and variate attention. The 209M checkpoint uses 24 layers at an embedding width of 768, with each patch covering 32 time steps. Context reaches 32,768 steps, and a single pass forecasts up to 1,024 ahead.
Latency is the selling point. On an eight-thread server CPU, forecasting one day ahead from three weeks of 15-minute history took 10 ms at 5M parameters and 140 ms at 209M. Scaling flattens at the top, with the 123M model scoring 0.659 MASE against 0.658 for the full checkpoint.
Mistral reports that even the 28M model beats Chronos-2 on GIFT-Eval and TIME, despite Chronos-2 being about four times larger. All scores are zero-shot from single checkpoints, without ensembling, and come from Mistral’s own model card. The code is on GitHub with weights on Hugging Face.
