Mistral releases Timestral open weights in four sizes for probabilistic forecasting

Mistral has introduced Timestral, its first family of foundation models designed specifically for time-series forecasting rather than conversational text generation. Announced on October 15, the open-weight release comprises four checkpoints with 5 million, 28 million, 123 million and 209 million parameters. The models are intended to predict future values from historical measurements without task-specific training, including workloads in operations, energy, finance or sensor monitoring where uncertainty matters alongside a single forecast.

The family supports multivariate probabilistic predictions, past-only covariates and known-future covariates. Mistral's official model documentation describes context windows of up to 32,000 timesteps and quantile outputs, a different notion of context from the token window of a general-purpose language model. Mistral attributes the design to synthetic pretraining, covariate-aware attention and long-context learning. A forecast model's support for an input type does not itself establish accuracy or calibrated uncertainty for any particular enterprise dataset.

Mistral reports that the 209M model leads the public zero-shot comparisons it presents for GIFT-Eval, TIME and fev-bench. It also reports CPU inference examples on an eight-thread server: 10 milliseconds for the 5M variant and 140 milliseconds for the 209M variant on the publisher's one-day-ahead example. These are vendor-controlled benchmarks and workload-specific timings, not independent Open Model Weights performance tests or promises of comparable speed on other hardware.

The licensing distinction is significant. Mistral labels the downloadable models Modified MIT, not unrestricted standard MIT; its published license guidance describes extra commercial requirements for companies above a revenue threshold. For the registry, Timestral is four distinct checkpoint candidates with different deployment footprints. Each model needs its own pinned repository, actual weight-file evidence, license text and configuration before verified fields can be assigned. The release broadens the open-weight landscape into forecasting, but its benchmark ranking should not be turned into a universal model-quality score.