Mistral AI releases Timestral time series foundation model family

What happened

On 15 October 2026, Mistral AI announced the release of Timestral, its first family of time series foundation models (TSFM) [2]. A time series foundation model is a system trained on a large, diverse collection of data that can perform “zero-shot” forecasting, meaning it can predict patterns on new data it has never seen before [2].

The Timestral family includes four different model sizes based on the number of parameters they use: 5M, 28M, 123M, and 209M [1, 2]. According to Mistral, the largest 209M model ranks first among pretrained, zero-shot models on three public benchmarks: GIFT-Eval, TIME, and fev-bench [1, 2]. Mistral claims this model outperforms larger models like Google’s TimesFM-3 and Falcon-2.0 [1, 2]. Specifically, the 28M model is reported to beat Chronos-2, which is approximately four times larger, on the GIFT-Eval and TIME benchmarks [1, 2].

Technically, Timestral is a causal, patch-based transformer that uses time and variate attention [1]. It provides probabilistic, multivariate forecasts, meaning it can predict multiple related variables at once and provide a range of possible outcomes (21 quantiles per time step) to show uncertainty [1, 2]. The models can handle past-only data, known-future information, and categorical covariates, such as the day of the week [1, 2].

Efficiency is a key feature of this release. The models are designed to run on standard server CPUs without needing a GPU [1]. In tests using a standard server CPU with 8 threads to forecast one day ahead from three weeks of 15-minute history, the 5M model took 10 ms and the 209M model took 140 ms [1]. The 28M model is five times faster than the 209M version [1].

Who is affected

This release is aimed at developers and industries that rely on time series data, such as finance, energy, and supply chains [2]. For example, industrial plants can use these models to monitor thousands of sensors, forecasting equipment behavior by looking at how variables like temperature and pressure move together [2]. Finance teams can use them to forecast volumes and volatility across many accounts [2].

What to do next

Developers can install the package using the command `pip install git+ The software requires Python 3.10 or higher and torch 2.4 or higher [1]. Model weights can be downloaded from Hugging Face, with file sizes ranging from 21 MB for the 5M model to 490 MB for the 123M model [1]. The 209M model is currently in Public Preview [1]. For those needing to adapt the models to private data, Mistral offers fine-tuning through Mistral Forge [1, 2].