By Hamza Faisal

Mistral AI has released Timestral, its first family of time-series foundation models — and the French lab says the largest of the four models beats Google's TimesFM-3 on every major forecasting benchmark, no GPU required.

Released on October 15, the Timestral family spans four dense models sized at 5M, 28M, 123M and 209M parameters. All four share the same patch-based transformer architecture and training recipe, so developers can swap sizes without rewriting code.

A forecaster that reads numbers like a language model

Classical forecasting tools such as ARIMA fit one series at a time, and deep models like PatchTST usually train from scratch on each dataset. Timestral works differently: pretrained on a huge mix of data — mostly synthetic, according to Mistral — it can forecast series it has never seen before, a capability known as zero-shot forecasting.

Each model outputs 21 quantiles per time step, giving probabilistic forecasts — a range of likely outcomes rather than a single number. It forecasts multivariate targets jointly, so related signals like a pump's pressure and temperature are predicted together, and it accepts past-only and known-future covariates plus categorical inputs such as the day of the week or whether a machine is running.

The largest checkpoint, Timestral-209M, ranks first among pretrained, zero-shot models on GIFT-Eval, TIME and fev-bench, per Mistral's reported results: on fev-bench's SQL skill score it posts 50.39 against TimesFM-3's 48.99, and on GIFT-Eval it edges t0-gamma 0.658 to 0.661. Mistral also claims its 28M model outperforms Chronos-2 — which carries roughly four times its parameters — on GIFT-Eval and TIME.

No GPU? No problem.

The headline feature for infrastructure teams: Timestral runs on a standard server CPU. Mistral reports that forecasting one day ahead from three weeks of 15-minute historical data takes about 140 milliseconds on an eight-thread CPU for the 209M model, and roughly 10 milliseconds for the smallest.

Weights and a lightweight inference library are open — the code ships under Apache 2.0. The weights, however, carry Mistral's "Modified MIT" license, and the modification matters: companies whose global monthly revenue exceeded $20 million in the preceding month are not authorized to use them.

Mistral says most of the training data is synthetic — univariate processes with trend, seasonality and noise, combined into multivariate series with nonlinear dependencies — and that a full technical report is coming soon.

Why it matters

Time-series data runs the business world: factories track vibration and temperature, finance teams watch cash flows, IT teams monitor server traffic. Accurate forecasts drive decisions from spare-parts orders to power planning — and a model that runs on ordinary hardware lowers the cost of entry dramatically.

The release also pushes Mistral beyond chatbots into specialized enterprise tooling, a space Google has contested with TimesFM. Whether Timestral's benchmark leads hold up under independent testing will decide if the French lab's forecaster becomes the new default for the numbers that keep businesses running.

By Hamza Faisal