French AI company Mistral released Timestral on October 15, 2026. It is the company’s first family of time-series foundation models, built to forecast how numbers change over time.
The largest Timestral model has 209 million parameters. Mistral says it ranks first on GIFT-Eval, TIME and fev-bench, the three main public benchmarks for forecasting.
According to the company, the model beat the previous state of the art while using only two-thirds as many parameters. Mistral has submitted its results to the GIFT-Eval and TIME leaderboards.
Four Sizes, Built to Run on CPUs
Timestral comes in four sizes, ranging from 5 million to 209 million parameters. Mistral says each size delivers leading accuracy for its size, placing the family on what it calls the Pareto front of accuracy versus model size.
The company says the models can run efficiently on CPUs. These are the standard processors found in most computers.
Timestral produces probabilistic forecasts, meaning it gives a range of possible outcomes rather than a single number. It also handles multivariate data, so it can forecast several related series at once.
The model supports covariates, which are outside factors that affect a forecast. These include past-only data and information known in advance about the future, such as a planned production schedule.
Mistral says the model combines synthetic pre-training, covariate-aware attention and long-context learning. Users who want to adapt it to a specific field can fine-tune it through Mistral Forge.
How Finance and Industry Teams Use Forecasts
Mistral says time series is one of the most common types of data in business. It ranges from sensor readings in factories to performance metrics in computer systems.
According to the company, finance teams forecast flows, volumes and volatility. They use these forecasts to guide positions, hedging and treasury decisions across thousands of accounts.
Industrial plants use forecasts to compare sensor data against production schedules. They also use them to spot equipment drift before a failure and to plan spare parts across thousands of assets.
Classical tools such as ARIMA and linear regression are still widely used. These models are usually fitted to one series at a time and learn only from that series’ own history.
Deep learning models such as PatchTST can learn more complex patterns across many related series. However, they are usually trained from scratch on each dataset, so what they learn does not carry over to new areas.
Foundation models take a different approach by training on a large and varied set of data first. Mistral says this lets Timestral make zero-shot forecasts on data it has never seen, even where data is scarce.
Mistral tested Timestral on all three benchmarks against other pre-trained, zero-shot models. The company says it ranked first across all five metrics measured.
These results came from a single model checkpoint. Mistral says it did not use ensembling, a method that combines several models to raise scores.
