Mistral's Timestral Beats Forecasting Leaders on CPUs With Just 209M Parameters

Mistral AI has released Timestral, a family of open-weight foundation models for time-series forecasting. The release includes four checkpoints, public inference code, and weights hosted on Hugging Face. Every model supports CPU-only inference, giving developers a path to zero-shot forecasting without operating GPU infrastructure or training a separate model for each dataset.

According to Mistral’s results, the 209M-parameter checkpoint ranks first on GIFT-Eval, TIME, and fev-bench. The company says it surpasses the previous state of the art with roughly two-thirds as many parameters. Those results are self-reported, so production decisions still require independent testing on representative data.

Four sizes, laptop-scale weights

All four checkpoints fit in typical laptop or server memory. Actual runtime consumption will exceed the weight sizes because inference also allocates memory for activations, input arrays, and framework overhead.

• Zero-shot forecasting: Generates predictions without task-specific training.
• Multivariate forecasting: Models related target variables jointly.
• Probabilistic output: Returns 21 quantiles for each predicted time step.
• Long context: Accepts as many as 32,768 historical time steps.
• Covariates: Supports past-only, known-future, continuous, and categorical signals.
• Missing data: Accepts missing observations represented as NaN.

Mistral reports CPU inference latency of about 10 ms for the 5M model and 140 ms for the 209M model on an eight-thread server CPU. Latency will vary with batch size, context length, forecast horizon, number of variables, numerical precision, and host hardware, so those figures should not be treated as end-to-end service guarantees.