Mistral Timestral: 209M-parameter forecasting model runs on CPU

Mistral released Timestral on 15 October 2026, its first family of time-series foundation models. These are pre-trained models that forecast numeric series, such as sensor readings, demand or capacity metrics, on data they have not seen before. Everything below is Mistral's own account; the company has said a technical report with full results will follow.

What Mistral released

Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters. Mistral lists a context length of 32,768 time steps and a licence of modified MIT. The models forecast several related series at once and return 21 quantiles rather than one number, so a team can see how wide the uncertainty is and set alert thresholds accordingly.

They also accept covariates, meaning other series that help explain the target. Mistral says that covers past-only inputs and known-future ones, such as production schedules, planned shutdowns and weather forecasts, plus categorical inputs such as day of the week or machine state (running, idle, under maintenance).

What Mistral reports on accuracy and speed

Mistral says its 209M model ranks first on GIFT-Eval, TIME and fev-bench among pre-trained, zero-shot models, across all five metrics, from a single checkpoint without ensembling, and that it beats the previous state of the art with two-thirds of the parameters. It says the 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME, and that the 209M model beats TimesFM-3 and Falcon-2.0. Mistral says it has submitted results to the GIFT-Eval and TIME leaderboards; the comparisons shown on the page are Mistral's, not an independent evaluation.

On cost, Mistral says no GPU is needed at inference. On a standard server CPU with 8 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. It adds that the 123M model comes within 0.2% of the 209M score at three quarters of the latency.

How it is meant to be used

Mistral describes a plant that runs Timestral continuously on sensor streams to flag anomalies. An agent then compares a flagged pump against past failures, starts sub-agents for root-cause analysis using maintenance records and shift logs, and drafts a diagnosis for the plant manager. That is an illustration from the announcement, not a reported deployment.

For adaptation, Mistral points to its Forge training pipeline, which it says supports fine-tuning on proprietary data and can be deployed on a customer's own hardware, including air-gapped environments. It also lists a GitHub repository for running the model on your own data.

What is not stated

The announcement gives no price for Forge, no results on any customer's data, and no head-to-head against the classical methods, such as ARIMA, that many teams still use. Mistral says most pre-training data was synthetic, then post-training mixed real and synthetic data, but the technical report that would describe the data in detail has not been published. A reasonable test is a backtest of the 28M or 123M model against your current forecaster on a few of your own series, scored on the same metric you already report to finance or operations.