Mistral released Timestral, a 209M time-series model that beats larger ones

The forecasting model running on a factory's sensors is almost always one somebody on the team trained from scratch, and it forgets everything the moment the data moves on. Mistral AI is proposing the same shift for time series that language models already went through: a pretrained foundation model you can use zero-shot and fine-tune later.

Mistral released Timestral, a family of four open time-series foundation models at 5M, 28M, 123M and 209M parameters. Code is Apache 2.0 at mistralai/timestral on GitHub, and the weights carry a Modified MIT licence on Hugging Face.

What the ranking claim says

Mistral says the 209M model ranks first among pretrained and zero-shot entries on three public forecasting benchmarks: GIFT-Eval, TIME across 98 tasks and fev-bench across 100 tasks. The company says all four model sizes sit on the Pareto front of accuracy against size. In Mistral's own comparison, the 28M model beats Chronos-2, a model about four times larger, on GIFT-Eval and TIME, and the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. These are the vendor's own figures from a single checkpoint without ensembling, and results have been submitted to the GIFT-Eval and TIME leaderboards.

Runs on a CPU

Timestral does not need a GPU for inference. Mistral reports that on a server CPU with eight threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model. The 28M model is five times faster than the 209M, and the 123M comes within 0.2% of the 209M score at three quarters of its latency. For teams already running forecasts inside monitoring or planning pipelines, that is the difference between a line in a Python job and a request for GPU time.

What the model takes as input

The family handles multivariate targets jointly rather than one series at a time, so related signals such as a pump's vibration, temperature and pressure share information. It accepts past-only covariates, known-future covariates like a production schedule or a weather forecast, and categorical covariates like the day of the week or a machine's state. Context length is up to 32,768 steps. The output is probabilistic: 21 quantiles at each future step, which lets a team set alert thresholds against the width of the forecast rather than its centre.

Training recipe, in Mistral's words

Mistral says most of the training data was synthetic. The team generated univariate processes with trends, seasonality and varied noise, then combined them into multivariate series with nonlinear dependencies. Timestral is first pretrained on synthetic data with context 8,192, then post-trained on a mix of real and synthetic data at four times that length. The loss is pinball loss over 21 quantiles. Mistral also describes a technique it calls Decoupled Horizon Forecasting, which keeps the full history in the sequence while learned forecast tokens read it, so every observed value serves as both target and context. A technical report is promised.

Timestral gives a team whose current pipeline is one ARIMA per sensor or one LSTM per dataset a pretrained starting point that runs on the hardware they already have, with a fine-tuning path through Mistral Forge on their own data if the public results do not transfer to the domain.

Source

Timestral on mistral.ai and mistralai/timestral on GitHub.