Mistral releases Timestral time-series forecasting models

Mistral has released Timestral, a family of four time-series foundation models with 5M, 28M, 123M and 209M parameters. All four use the same architecture and training recipe. The models produce probabilistic, multivariate forecasts and accept both past-only and known-future covariates, including categorical inputs such as day of week or machine state.

Mistral says its 209M model ranks first among pre-trained, zero-shot models on all five metrics across the field's three main public benchmarks: GIFT-Eval, TIME and fev-bench. According to the company, those results come from a single checkpoint without ensembling, and the model beats the previous state of the art with about two-thirds of its parameters. Mistral also claims the 28M model beats Chronos-2, which is roughly four times larger, on GIFT-Eval and TIME, and that the 209M model outperforms TimesFM-3 and Falcon-2.0. These results are self-reported. Mistral says it has submitted them to the GIFT-Eval and TIME leaderboards, but they have not yet been confirmed there.

Inference does not need a GPU. On an 8-thread server CPU, forecasting one day ahead from three weeks of 15-minute data takes 10 ms with the 5M model and 140 ms with the 209M model. The 123M model scores within 0.2% of the 209M model at three-quarters of its latency.

Mistral describes training that combines synthetic pre-training, covariate-aware attention and long-context learning. Customers can fine-tune the models for their own domains through Mistral Forge. The company pitches Timestral for industrial and financial forecasting, either inside existing monitoring pipelines or alongside its language models in agentic workflows. The weights are reportedly available under a modified MIT licence. A technical report has not yet been published.