Mistral AI released Timestral on October 15, 2026. It is the company’s first family of time-series forecasting models. The four models, ranging from 5 million to 209 million parameters, can predict multiple related data streams and provide a range of possible outcomes.
Mistral says its 209 million parameter model ranks first across five metrics on the GIFT-Eval, TIME and fev-bench public benchmarks among pre-trained models tested without additional training. The company says the results come from a single model checkpoint. It has submitted results to the GIFT-Eval and TIME leaderboards.
The models can run on standard CPUs
All four models can run on a CPU, according to Mistral. On a server CPU using eight threads, the company reports that forecasting one day ahead from three weeks of readings taken every 15 minutes takes 10 milliseconds with the smallest model and 140 milliseconds with the largest.
Timestral can use past observations alongside information known in advance, such as a production schedule or weather forecast. It also accepts categorical inputs, such as a machine’s operating state. The models support a context of 32,768 time steps and produce forecasts across 21 quantiles to show uncertainty.
Organizations can fine-tune models on client hardware
Mistral says organizations can use Timestral through Mistral Forge for fine-tuning on their own data and for changes such as adding inputs specific to a domain. The training pipeline can run on clients’ hardware, including in environments disconnected from the internet. The models carry a Modified MIT license.
Mistral has published a GitHub repository for running Timestral on users’ data. The company says a technical report with the full results will follow.
