Skip to content
Neural Daily

ModelsNews

Mistral Timestral: time-series forecasting on CPU

Mistral released Timestral, its first family of time-series foundation models. The largest variant has 209M parameters and, according to Mistral, ranks first on all three main public forecasting benchmarks. Inference does not need a GPU.

By: Mikola Máté Patrik · · 2 min read

Mistral AI Timestral cover image: pixelated orange and blue hourglass on a dark background
Image: Mistral AI · Press kit · Official article image · Source

What Mistral released

Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters. It produces probabilistic, multivariate forecasts and supports both past-only and known-future covariates, such as a production schedule or a weather forecast (Mistral). Context length is 32,768 time steps and the output is 21 quantiles. The licence is a modified MIT.

Mistral AI diagram: Timestral model overview with patch embedding, transformer backbone and quantile head
Mistral AI · Press kit · Vendor diagram from the launch post · Source

Source: Mistral AI. Timestral architecture: patch embedding, a transformer backbone with temporal and variate attention, then a quantile head.

What the benchmarks show

Mistral says the 209M model ranks first among pre-trained, zero-shot models on all five metrics across GIFT-Eval, TIME and fev-bench, from a single checkpoint without ensembling. The 28M model beats Chronos-2, about four times larger, on GIFT-Eval and TIME, and the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. Results have been submitted to the GIFT-Eval and TIME leaderboards. The full technical report is still to come.

Why CPU inference matters

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 28M model is five times faster than the largest, and the 123M model comes within 0.2% of the 209M score at three quarters of its latency. That makes it easy to drop into an existing monitoring pipeline without a GPU fleet.

Where it fits

Mistral's example is a plant's sensor data: a pump's vibration, temperature and pressure move together, and the model handles them jointly. Quantiles expose the uncertainty, so alert thresholds can reflect risk. The same logic applies in finance, energy and capacity planning. Mistral also pitches pairing it with its language models in agentic workflows, for example automated root cause analysis after an anomaly.

How to try it

There is a GitHub repository for running Timestral on your own data (GitHub). Fine-tuning is available through Mistral Forge, which supports on-prem, VPC and air-gapped deployment. If you forecast time series today, benchmark it against your current ARIMA or Chronos setup on your own backtests.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles